All notable changes to TinyBooth Sound Studio.
The format follows Keep a Changelog; this project tracks Semantic Versioning loosely (the v0.x series treats minor bumps as feature releases, patch bumps as fixes / polish).
(Nothing yet — known issues all resolved as of v0.4.23.)
The in-app updater is rebuilt around a signed-release trust chain:
- Every release now ships a signed manifest (
RELEASE.manifest+.sig): the version and the SHA-256 of each asset, signed with a release key whose public half is built into the app itself. The updater verifies the signature, the version, and the installer hash before anything runs elevated — a compromised download, a tampered release, or a replayed old version is refused, not risked. - Fail closed. A release without a valid signed manifest gets no install button and no silent fallback — the previous updater only checked integrity when a checksum happened to be published.
- The installer now launches through the Windows shell API directly — no PowerShell command line anywhere in the path — and downloads stream to disk with real progress (MB counts in the status bar).
- CI refuses to publish a release it cannot sign.
The updater lives in a new reusable selfupdate crate, built to be
retrofitted into every installable app (starting with the skeleton
this app's tooling came from). Not claimed: Authenticode — that
requires a code-signing certificate; the config already carries the
slot for when one exists.
- A Start-menu shortcut for the installed app (retrofit from the shared app skeleton).
- 🌐 Sample Library. Download free instrument packs from inside the Tracker — Philharmonia Orchestra (strings / brass / woodwind / percussion) and the CC0 Versilian Community Sample Library. Licenses are shown before download; packs come from their sources on demand and nothing is bundled with TinyBooth. Verified against the real packs: the Brass zip alone yields 2,940 per-note recordings, every one auto-mapped to its pitch from the filename; unpitched percussion is honestly reported rather than misfiled.
- Multisample instruments. One pack click builds an instrument with a zone per note — playing C-3 uses the C-3 recording, not a five-octave stretch of one sample. Existing single-sample instruments and songs load and render exactly as before (pinned by test).
- Five-octave piano editor. Each instrument shows a C-2…B-6 piano: green keys hold a real recording, others pitch-stretch the nearest zone; clicking a key auditions it through the real engine (filter, gain, loop and all).
- Per-key wave editor. The selected key opens its zone's waveform with the familiar wave-lane language — and dragging a selection sets that zone's trim directly.
- Demo songs. A Demos menu loads bundled public-domain classics — Ode to Joy (Beethoven, 1824), Frère Jacques (trad.), and Korobeiniki (trad., 1861 — yes, that one) — as ready-made patterns that play through whatever instrument you've built.
Six findings from the full-codebase audit, ordered by severity:
- Self-update is no longer blind trust. Releases now publish a SHA-256 sidecar next to the installer, and the in-app updater verifies the download against it before anything runs elevated — a mismatch or an implausibly small download is refused outright. The release version string is also sanitized before it touches a filename or command line.
- Re-migrating a folder can no longer destroy an existing
.tib. Migration now refuses if the target exists (with a clear message), runs as a single transaction, and removes the partial file on failure — a crash mid-migration used to leave a half-populated container that opened without complaint. - The folder-project manifest is saved atomically (temp + fsync +
rename) with a
.bakof the previous version — a crash or full disk mid-save used to truncate the only copy of all mix state. - Analysis no longer silently zeroes on hi-res files. Telemetry and the import coherence check read samples as 16-bit unconditionally, so a float or 24-bit WAV produced empty/garbage analysis with no error. Both now decode through the canonical path; coherence also handles any channel count instead of refusing >2.
- Exporting projects with 24/32-bit sources no longer wrecks the balance. The export decoder scaled all int samples as 16-bit, so a 24-bit stem entered the mix ~256× hot and the limiter crushed everything else to compensate.
- A corrupt settings file is preserved as
config.json.corruptfor recovery instead of being silently replaced with defaults.
A MadTracker-lineage tracker lands as a new Tracker tab:
- Pattern editor — vertical grid (rows down, tracks across), hex row numbers, bar shading, cursor; click to focus, then play notes on the keyboard (FastTracker layout: ZXCV… lower octave, QWER… upper; F9/F10 octave, arrows/PgUp/PgDn/Home/End navigation, Delete clears, edit-step auto-advance). The grid is the first tenant of the v0.4.82 keyboard interface's editor scope — it owns the keys while focused, and Ctrl+S / F1 still work.
- Instruments — add any audio file as a sample (decoded in the background); per-instrument base note, gain, loop off/forward/ ping-pong, cut/ring (NNA), and a resonant low-pass filter.
- Engine — classic tick timing (speed + BPM), variable-rate pitch with authentic aliasing, and the classic effect set (arpeggio, slides, vibrato, sample offset, volume slide, pattern break, speed/BPM, note cut, note delay) — all pinned by 13 frame-exact tests.
- Transport — ▶ loops the pattern (edits re-render at the loop boundary); ⤓ Bake as stem adds the loop to the project as a track; Export WAV… writes it out. Tracker playback participates in the app-wide one-audible-thing rule.
- Songs persist automatically per project.
v1 fences (follow-ups tracked in the FR): volume/pan/effect columns are engine-supported but not yet keyboard-editable (notes only), one pattern (order chains play but no order editor), no drum-grid view yet, instrument picker is file-based (recordings-take picker to come).
- Scrubbing. The Mix playhead is no longer display-only: click any lane or the new timeline ruler to seek, or drag the playhead itself to scrub. (Wires up the seek machinery that had sat dormant as "Phase 3" since the player was written.)
- Region highlight. Click-drag on a lane selects a time range on that track; drag on the ruler selects across all tracks. The band draws on every lane it covers.
- ✂ Clip. With a highlight active, a toolbar offers Clip to WAV… and Clip to new track: the range renders through the full playback chain — enabled corrections, faders, automation, polarity, and master values — on a background thread. Renders process from t = 0 so filter state at the range boundary matches a full render sample-for-sample (verified by test). A single-track clip carries only that track's signal (verified by test), still through master.
The export/bounce mixdown historically dropped polarity flip (Ø)
— an exported mix could differ audibly from playback for any project
using polarity correction (anti-phase stems failing to cancel, or
cancelling when playback didn't). Polarity is now applied in the same
place playback applies it, for exports, bounces, and clips alike. If a
project uses Ø, re-Bounce to refresh its cached mix.
The app now has one keyboard-ownership model instead of ad-hoc key checks: text fields > open modals > editor widgets > global shortcuts, resolved each frame from a single declarative binding table (which a future shortcut-help overlay and rebinding UI will read too). First bindings:
- Space — play/pause the current audible thing (a playing recording preview stops; otherwise the mixer toggles).
- Ctrl+S — save the project.
- F1 — toggle the manual (migrated from the old one-off check).
- Esc — close the topmost open dialog. The folder→
.tibmigration prompt deliberately requires an explicit click.
The "editor scope" tier exists for widgets that own the keyboard while focused — designed for the upcoming tracker's FastTracker-style pattern entry, which claims keys without stealing Ctrl+S/F1. Precedence rules are pure and unit-tested, including that shortcut consumption can never eat a text field's or an editor's keys.
Reshaped per direct user feedback ("lets stop using mix for this") after the ▶-to-mixer flow failed once more — an async player-build race could drop the solo request, so pressing ▶ on one take played all of them. Rather than patch a fourth bug in that flow, the flow is retired:
- ▶/■ on each take, right in the listing. Plays through its own lightweight session — the Mix tab is not involved. While playing, the waveform thumbnail becomes a transport: an amber playhead tracks the position and clicking or dragging scrubs. When stopped, dragging still selects a region for export, as before.
- One audible thing at a time. Starting a take silences the Mix player and the Crossfade/Album previews — and starting any of those silences the take.
- ⇪ Integrate into project. The new "work on this take" gesture:
copies the take into the currently open project as a real track —
where the full cleaning toolset (corrections, telemetry, trim,
mixing) applies. Works for both folder and
.tibprojects; refuses a sample-rate mismatch with a clear message instead of landing a track the mixer would skip. - The v0.4.80 take-detail view is removed along with the detour it patched.
A Monitor toggle next to ⏺ Record opens the input stream with no WAV writer: waveform, spectrum, and meters run through the exact recording-tone chain a take would use, so base levels are visible before committing to a recording. Starting a take hands the device over automatically. Verified on hardware: frames flow through the chain, nothing is written to disk.
Clicking ▶ on a take in the Record tab's recordings list now lands in a detail view of that one take — its lane and its console strip (plus the master strip), under a "Take detail" banner — instead of the full recordings mixer with every take visible. "⊞ Show all takes" widens back out to the whole project and un-solos, so the other lanes come back audible. Follows directly from user feedback on v0.4.79, where ▶ correctly played the selected take but still showed all of them.
Both reported minutes after v0.4.78 restored recording. Two independent defects:
- ▶ could play a different take than the one clicked. The mixer refuses to load tracks whose length differs from the first one it loads — a safety rule for stem projects (where every stem must line up) that is exactly wrong for the recordings list, where takes are naturally all different lengths. With three takes only the first actually loaded, and the play button's track lookup was positional, so clicking ▶ on one take solo'd — and played — another. Recording projects now load all takes regardless of length (shorter ones simply end early), every lane carries a durable link back to its project row, and all fader / polarity / automation / correction write-backs go through that link. Clicking ▶ on a take that genuinely couldn't load now says so instead of silently playing something else.
- Waveform thumbnails were permanently blank for new takes. The list renders while a take is still recording, and a thumbnail computed from the not-yet-finished WAV (which reads as zero-length until the recorder closes it) was cached forever. Thumbnails now refresh automatically when the file changes, and the take currently being recorded is never thumbnailed mid-write.
Pressing ⏺ Record silently failed on any recordings project that already had at least one take (user-reported with a Yeti Stereo Microphone: "nothing happens when I play record"). Two defects stacked:
- Format-blind device configuration. Once a project has a take, the recorder pins new takes to that sample rate — and the rate-matching path picked the first device configuration covering the rate, whatever its sample format. On Windows the Yeti advertises an 8-bit configuration first, which the recorder can't stream, so every take after the first failed with "unsupported sample format U8". (This was also the long-suspected FR-0008 "repeat-take race" — no race: first-take-works/next-take-fails was the rate pin kicking in.) Configuration selection now considers only streamable formats, preferring float, then 16-bit, with channel count as the tiebreaker so Stereo mode keeps working.
- The failure was invisible. The error went only to the status bar at the very bottom of the window, so the click looked like a no-op. Record failures now also appear in red directly under the Record button.
Verified on the reporting hardware via a new (ignored, opt-in) probe test that runs three consecutive takes through both configuration paths against the real device: 3/3 takes start in ~30 ms at 48 kHz.
A five-agent fresh-eyes audit of the entire workspace (~35,400 lines) confirmed and fixed four defects, three of them data-loss:
- A migrated
.tibproject lost its Suno mixdown on the first save. Migration stores the mixdown under a reserved track that is never part of the in-memory project, and the save-time prune treated "absent from the project" as "deleted by the user" — the SQL cascade then removed the mixdown's audio for good. Pruning now only considers real mix lanes, and a new regression test (which failed before the fix) pins the behaviour: migrate → load → save → mixdown still readable. - File → Quit skipped shutdown entirely. It terminated the process outright, bypassing the exit hook — so a recording in progress never had its WAV header finalised (an unreadable take), and settings went unsaved. Quit now closes through the same clean path as the window ✕.
- Mix-tab track faders were never saved. Fader moves reached the live player but were never written back to the project, so they vanished on any player rebuild and were ignored by Save and Export — silently discarding balance work unless automation was armed. The fader now writes through and marks the project dirty, exactly as the master fader and polarity button always did.
- 24-bit WAVs played back as a square wave. Four separate decode paths (playback, recording thumbnails, crossfade, album) all clamped raw 24/32-bit samples into 16-bit range instead of scaling them — anything louder than about −48 dBFS pinned to full scale. All four now share one canonical decoder with correct bit-depth scaling, verified by a test that a −20 dBFS 24-bit sine decodes to −20 dBFS, not to full scale.
- First tranche of the audit's cohesion plan: the four copy-pasted WAV
decode ladders and two stereo-expansion loops collapsed into
audiodecode::decode_wav_i16/wav_i16_to_stereo_f32(~130 duplicated lines removed). - The audit's full findings — per-feature fit-for-purpose verdicts, the ranked consolidation plan, and production-completeness definitions — are recorded outside the repo; open items land in future releases.
v0.4.75 shipped with a known limitation: on dense, real-world mixes the
analyser over-segmented badly and labelled almost everything a seventh. On a
four-minute rock track it produced 494 chord spans — one every half
second — with a histogram of Dm7×61 Dmaj7×59 D7×59: major, minor and
dominant on the same root in near-equal numbers, which cannot all be right.
Three separate defects were behind that. Each was found by running real audio through the pipeline, and each fix is measured against it rather than assumed.
Tempo estimation — two bugs.
- The integer frame-lag was rounded to nearest, so the fast edge of the 60–180 BPM search band became lag 14 at 43.07 fps — 184.6 BPM, a tempo outside the range being searched. Lag bounds now round outward (ceiling at the fast end, floor at the slow end) and the result is clamped.
- The candidate score divided the autocorrelation by the lag, described as a preference for slower tempi. It did the opposite: the overlap window already shrinks as lag grows, so dividing by lag compounded a bias toward short lags rather than countering it — which is why real music pinned itself to the fastest available bucket. The score is now normalised by the overlap count, with a log-Gaussian tempo prior that is symmetric in octaves, so half-time and double-time are penalised equally.
- Added parabolic sub-frame refinement of the winning lag. Integer lags quantise about 12 BPM apart at the fast end (lag 14/15/16 land on 184.6/172.3/161.5), which is far coarser than the tempo itself.
On the test track: 184.6 → 126.8 BPM, inside the band.
Chords are now decoded as a sequence, not beat by beat.
Each beat used to be labelled independently, so ordinary chroma wobble on a dense mix flipped the winning template constantly. A chord is a held object, so the whole track is now Viterbi-decoded across the 72 chords plus a "no chord" state, with a self-transition prior: holding the current chord is free, changing costs a fixed penalty, and a candidate must clearly beat the incumbent to displace it.
The penalty was swept against the real track rather than guessed (the table lives in the source next to the constant). It is set toward the over-segmenting end on purpose: the editable grid can merge neighbouring spans trivially, whereas a chord change the decoder never emitted cannot be recovered by editing at all.
Sevenths no longer win on arithmetic alone.
L2-normalised binary chord templates are not scale-fair across qualities: a four-note seventh spreads its unit norm across four bins (0.5 each) while a triad spreads it across three (0.577 each). On a dense chroma — where every pitch class carries some energy — the seventh simply captures more of it and wins the cosine match regardless of what is actually being played. Confirmed by neutralising the new prior and re-running: smoothing alone yielded a histogram of nothing but sevenths. Chord qualities now carry a prior, so an extension has to earn its extra note.
Net effect on the same four-minute track: 494 spans of
self-contradictory sevenths → 82 spans reading
D×16 A×9 G×7 Bm×7 F#×6 — a key-coherent I–V–vi–IV set in D with F# as
the dominant of B minor.
TBSS_CHORD_FIXTURE=<audio file> runs any track through the full pipeline
and reports tempo, span count, mean confidence, a chord histogram, the
detected verb, and a check that every chord resolved to a voicing in the
database; add TBSS_CHORD_RENDER=1 to mux the video too. It is #[ignore]d
and skips cleanly when unset, so ordinary test runs are unaffected. This is
what surfaced every defect above, and the v0.4.75 ffmpeg regression before
them — synthesised fixtures had hidden all of it.
Chord recognition on full-band material is much better but not solved. Verb (repeating-progression) detection still finds nothing on real songs, and while the tempo is now stable and in-band, its octave choice is unverified — a detected 126.8 BPM may be double the musical tempo. The editable grid remains the mitigation, and remains first-class.
Chord-video rendering was broken in v0.4.74 for anyone on a recent
ffmpeg, including the build bundled with the installer itself. ffmpeg
removed the -vsync option between its 2026-06-15 and 2026-08-17 builds,
so the render aborted immediately with:
Unrecognized option 'vsync'.
Error splitting the argument list: Option not found
Replaced with -fps_mode cfr, which has existed since ffmpeg 5.1 and works
on both sides of the removal — verified against three builds (N-126188
rejected -vsync; N-125049 and 6.0 accepted both).
This slipped through v0.4.74's end-to-end test because the test resolved an
older ffmpeg on the build machine than the one shipped in the installer. It
was caught by running a real song through the pipeline, and that harness is
now part of the suite: set TBSS_CHORD_FIXTURE to any audio file to get
tempo, span count, confidence, a chord histogram and a database-resolution
check (plus TBSS_CHORD_RENDER=1 to mux the video). It is #[ignore]d and
skips cleanly when unset.
Chord recognition accuracy on dense, real-world mixes is still weak — expect over-segmentation and unstable seventh/triad labelling on full-band material. The editable grid is the mitigation for now; an accuracy pass (tempo-band fix, temporal smoothing, template rebalancing) is in progress.
(Addressed in 0.4.76.)
A new Chords tab turns a song into that same song with a synced guitar fretboard-diagram video track over the original, untouched audio — one chord at a time, standard EADGBE tuning.
Load a song → Analyze → review and correct the detected progression → Render video…. Accepts wav, mp3, flac, m4a, aac, ogg, opus, wma.
- Chord analyser (
chordgrid.rs) — STFT (4096 / hop 1024, Hann) → spectral-flux onset envelope → autocorrelation tempo (60–180 BPM) → phase-locked beat grid → beat-synchronous chroma → cosine match against 6 qualities × 12 roots, plus verb (repeating-progression) detection. Every beat carries a confidence. - Generative chord database (
chorddb.rs) — rather than hand-entering shapes, voicings are generated: enumerate playable fret combinations in a moving neck window, keep only those whose sounding pitch-classes spell the target chord (root in bass, no foreign notes), assign an ergonomic fingering, then score and rank. 2,287 ranked voicings across 16 qualities × 12 roots, every one pitch-class-correct by construction. The 5th is modelled as optional, so the idiomatic open C7 (x32310, no 5th) is a first-class voicing rather than a special case. 44 hand-verified shapes act as the golden set — pinned to rank 0 so common chords render their recognisable open/barre grips, and each doubling as a correctness anchor in the tests. - Editable chord grid — the point, not a nicety. Full-mix recognition will miss borrowed and extended chords, so weak detections are flagged ⚠, any span's root/quality is editable (or settable to N.C.), and an edit re-resolves its voicing through the same path the analyser uses — a corrected chart and a detected one render identically.
- Voicing resolver (
chordvoice.rs) — collapses per-beat cells into spans and threads the previous fret position forward, so the diagrams don't leap up and down the neck between chords. - Fretboard renderer (
fretboard.rs) — pure software rasteriser, no GPU or font dependency, so output is deterministic. Horizontal neck, numbered fingers 1–4, open/muted markers, barre bars, and a left-handed toggle that mirrors at draw time (stored shapes stay canonical right-handed). One renderer feeds both the on-screen preview and the video frames. - Video build + mux (
chordvideo.rs) — one diagram per chord span via ffmpeg's concat demuxer (which encodes the beat timings directly instead of duplicating thousands of identical frames), H.264 video, and the audio stream-copied so it is never silently re-encoded. The H.264 encoder is probed rather than hardcoded, so builds withoutlibx264(which uselibopenh264or a hardware encoder) still work. - Options: frame rate, left-handed diagrams, and an opt-in "Re-encode audio (AAC)" for players that won't open PCM-in-MP4. It defaults off — your audio stays bit-exact unless you ask.
Analysis and rendering both run on a background thread, so neither blocks the UI.
- Visualizer chroma note labels were 9 semitones off. The pitch-class
mapping used an A-relative formula (
12·log2(f/440), putting A at class 0) while the labels were C-relative, so the chromagram and circle-of-fifths displays named the wrong notes. Corrected to the MIDI conventionround(69 + 12·log2(f/440)) mod 12in the shippedtbvizchroma, similarity, and hyperbolic modules.
Recovered the offline telemetry drum analyzer as a live streaming
detector inside tbviz, exposed it to every visualizer, and drew
counter-visuals from it. Works in TinyBooth today and TinyAmp later
(it runs off the shared FrameCtx, not a whole-track STFT).
tbviz::events::LickDetector— the same 5-band spectral-flux onset detection + adaptivemedian + k·MADpeak-pick + dominant-band drum classification (kick / snare / hat / tom / cymbal) astelemetry::classify_drum_events, reworked to run frame-by-frame off the rolling spectrum. Also estimates tempo (BPM) + beat phase from the kick/snare backbeat. Warmup-gated: produces nothing until it has enough history.FrameCtx.licks: Option<LickFrame>— the event track. TheOptionis the null-gate: a module physically cannot operate on a missing track (the type forcesif let Some(licks) = ...).Nonewhen detection is disabled, warming, or the audio is silent.- Across-the-board counter-visual — the shell draws class-tinted bubbles (rising, fading, per-instrument lanes) + a beat-pulse ring + BPM readout over whichever mode is active. When the track is absent it shows "no lick tracks present — listening…" instead of failing.
- Bespoke Mandala treatment — per the design brief ("if the visual is a spiral the licks show up as bubbles"): drum hits bloom as bubbles orbiting the mandala, flying outward along per-band spokes as they age.
- Two config toggles: Detect licks & beats and Lick/beat counter-visual.
3 new events::tests — silence → no events, low-band pulse train →
several Kick onsets (validates the recovered classifier fires + labels),
cold detector → None. tbviz suite 10 passing. fmt + clippy
--workspace --release --all-targets -D warnings clean under CI-pinned
Rust 1.95.0.
The visualizer (the VizModule trait, FrameCtx, all 20 modules, and the
shared helpers) moved out of src/ui/visualizer/ into a new workspace
library crate crates/tbviz. This is the keystone for TinyAmp — a
forthcoming portable player + heavyweight visualizer host — so both apps
share one engine with zero drift instead of copy-pasting modules.
No user-facing change: the 🌀 Visualizer behaves identically.
tbvizis host-agnostic — it knows nothing about projects, players, or files. Two couplings were severed: it carries its ownspectrum()DSP (soFrameCtx::build(samples, sr, time, dt)is self-contained) and inlines the coherence thresholds (mirroringtelemetry::COH_*).tbviz::show(state, ui, samples, sample_rate)replaces the old app-coupled entry point; aclose_requestedflag bridges the in-engine Close button back to the host.- TinyBooth's
src/ui/visualizer.rsis now a thin shim: it re-exportsVisualizerStateand feeds the live player tap totbviz::show. ci.ymlclippy + test gained--workspaceso the new member (incl.tbviz's hand-rolled TDA persistence test) is actually gated in CI.
Workspace builds; tbviz carries 7 tests (HSV, registry uniqueness,
centroid, Lorenz RK4 stability, TDA persistence circle/line); the binary
keeps 131. fmt + clippy --workspace --release --all-targets -D warnings
clean under the CI-pinned Rust 1.95.0. Files moved via git mv to
preserve history.
Completes the sound-visualization engine suite over the modular VizModule
architecture. Four new modes, all on the existing 2-D painter + FrameCtx
with no new dependencies:
- Topology (#15) — SW1PerS periodicity barcode. A sliding-window embedding of the signal is sphere-normalized into a point cloud; a hand-rolled Vietoris–Rips H₁ persistence (Z/2 boundary-matrix reduction) finds the persistent hole that periodicity creates. Renders the barcode + point cloud + a normalized score with a verdict (periodic / quasi-periodic / aperiodic). Recomputed every ~250 ms; the PH is unit-tested (a circle yields one long bar, a collapsed line none).
- Optical Flow (#16) — spectrogram tinted by instantaneous-frequency velocity (from the same derivative-of-window phase estimate as the reassigned mode). Vibrato shimmers as oscillating hue; glissando reads as a hue glide. Up-glide → blue, steady → green, down-glide → red.
- Saliency (#17, model-free) — Itti–Koch center-surround over the spectrum + a temporal-novelty term, surfacing what stands out without a trained model. The honest, offline-capable cousin of neural attention.
- Hyperbolic (#18) — agglomerative cluster tree of the chroma history, laid out on the Poincaré disk (depth → hyperbolic radius for exponential room), with Möbius pan/zoom by drag/scroll for infinite focus-plus-context navigation of song structure.
RAVE latent-walk (the neural-vocoder half of #17) is intentionally
omitted: it requires a pretrained model (tens of MB) and an ONNX/Torch
runtime that aren't available offline — bundling them would balloon the
69 MB installer and ship a tab that can't run without a user-supplied
model. The Saliency mode delivers #17's "what is the model/ear
attending to" question from first principles instead. Every other research
engine from docs/research/sound-visualization-engines.md now ships.
20 VizModules registered. New tda::tests (circle → one long H₁ bar;
line → none) validate the hand-rolled persistent homology. Suite
139 passing. fmt + clippy --release --all-targets -D warnings clean
under the CI-pinned Rust 1.95.0.
The visualizer was refactored from a 1,564-line monolith into a plugin system (a VizModule trait + a per-frame FrameCtx "TinyOutput contract" + a registry), then loaded with engines from the sound-visualization research (docs/research/sound-visualization-engines.md). 16 modes total, up from 5.
VizModuletrait — each visualization owns its params, persistent state,draw, andconfig_ui. Adding a new one is: createmodules/<name>.rs, re-export it, add one line todefault_modules().FrameCtx— built once per frame from the master-bus tap (stereo samples,sample_rate,time,dt, mono, magnitude spectrum, RMS, spectral centroid) and shared with the active module. The old code recomputed the FFT 3–4× per frame; now once.- Shared render helpers — a reused-texture
blit_image(2-D fields render as one texture upload, not 40krect_filledcalls), a perceptually-uniformmagmacolormap, andbin_hz. - Public surface preserved (
VisualizerState+show), soapp.rsneeded no changes.
- Spectrogram (#2/#4/#8/#12) — scrolling log-frequency, magma colormap, EMA auto-ranging. The baseline the app was missing.
- Reassigned (#3) — phase-reassigned spectrogram; 2× FFT (window + derivative-window), channelized instantaneous frequency relocates each bin. A/B toggle vs the raw STFT.
- Spectrum (#2) — log-frequency bars with EMA + falling peak-hold.
- Chroma (#5) — pitch-class energy on the circle of fifths + a resultant key-vector needle.
- Self-Similarity (#5) — cosine SSM of the chroma trajectory + Foote checkerboard novelty. Song form as geometry.
- Vectorscope (#10) — mid/side goniometer + running inter-channel correlation meter.
- Phase Portrait (#1/#6) — delay-coordinate embedding
x(t)vsx(t+τ), glowing orbit. - Recurrence (#6) — recurrence plot of the embedding with percentile-adaptive threshold.
- Timbre Map (#9) — online self-organizing map of a 7-D timbre feature; U-matrix + BMU comet. The autonomous-dimensioning showpiece.
- Particle Flow (#11) — curl-noise field driven by band energy + onsets (spectral flux), hue from centroid; frame-rate-correct advection.
- Health (#13) — scrolling spectral entropy / flatness / centroid / crest-factor ribbons. Surfaces over-compression, noisiness, monotony.
Lissajous, Mandala, Lorenz, Chladni, Onion Skin — now each a VizModule.
Five research engines are not in this build: TDA persistence barcode (#15, needs a persistent-homology crate), optical-flow LIC (#16) and hyperbolic form browser (#18) (no new deps — staged so this 11-module ship's build stays bulletproof), and neural saliency / Grad-CAM + RAVE latent-walk (#17, need ML runtimes + model files that aren't available offline). The color-honesty (#8), adaptive-display (#12), and rendering (#14) engines are embodied in the spectrogram's magma + auto-range + texture path rather than shipped as separate tabs.
16 modules compile; new default_modules_are_unique + centroid_of_empty_is_zero tests; the Lorenz RK4-stability + HSV tests moved with their code. Suite 137 passing. fmt + clippy --release --all-targets -D warnings clean under the CI-pinned Rust 1.95.0.
The biggest scope-shift since .tib itself: a sibling SQLite project format whose data model IS "arrangement of N stems", and an Album tab whose UI IS for editing that. After FR-0011 made .tib loadable as a single buffered stem, the natural next layer is a project file that composes N of those stems on a timeline — albums, mix tapes, long-form pieces, anything where the unit is a finished .tib and the compose-step is positioning + fades + gain.
SQLite v1 schema mirroring .tib's conventions (WAL, 16 KiB pages, foreign keys on). Three tables:
meta— name, created, schema_version.clips(id, ord, source_path, start_secs, fade_in_secs, fade_out_secs, gain_db)— one row per arrangement slot.source_pathis an absolute path to a.tibsource (relative-path resolution deferred to v0.4.53).mix_run(id=1, sample_rate, channels, frames, source_signature, created, audio BLOB)— structurally identical to.tib'smix_run. The bounced album's master goes here as a 16-bit WAV byte stream, so the Crossfade tab's existing.tibload path also accepts a.tbaas a stem with zero changes downstream. Albums-loaded-as-clips-in-another-album is recursive composition for free.
New Tab::Album variant + entry in the top tab strip:
- Header row: editable album name, Open / Save / Save As, dirty-indicator (
Save *), filename of the loaded.tba. - Top-down timeline visualisation: each clip rendered as a colored band at
[start, start+5s](the 5 s is a placeholder pending probe-clip-duration on add — landing in v0.4.53), with shaded fade-in / fade-out edges and the source filename baked into the band. - Clip list editor: scrollable grid with
#, source filename,Start (s),Fade in (s),Fade out (s),Gain (dB)DragValues, ▲▼ reorder buttons and ✖ remove.➕ Add Clip…button at the bottom opens an rfd dialog filtered to.tib. - Transport: ▶ Preview (renders + plays via the shared
CrossfadePreviewSession), ■ Stop, ⤓ Bounce (with✓ / ⚠ stale / —pip mirroring the Mix-tab Bounce button), Export… (any formatexport.rsalready supports).
Pure-function album::render(clips) -> AlbumMix. For each clip: equal-power fade-in (sin²(t·π/2)) over the first fade_in_secs · sr frames + equal-power fade-out over the last fade_out_secs · sr frames, multiplied by per-clip linear gain, summed into a single output buffer at the clip's start_secs. Soft-limit if any sample exceeds [-1, 1]. Adjacent clips' overlapping fades produce an emergent equal-power crossfade — no explicit transition object, the curves sum to constant power by construction.
All clips must share a sample rate (matches the constraint everywhere else in the app — no resampling). Mismatched-rate clips error clearly. Sources without a bounced mix_run error with the same "open and Bounce first" message the Crossfade tab uses.
Two new top-level modules + one new UI module + one new bridge module, mirroring the .tib / Project / tib_project / ui::mix layering:
tba.rs— SQLite storage (TbaDb,ClipRow,MixRunHeader, schema, CRUD + mix_run).album.rs— in-memoryAlbum+AlbumClip+ render DSP (load_clips,render,compute_mixrun_signature,encode_mix_to_wav_bytes).tba_album.rs—Album↔.tbashuttle (load_album,save_album).ui/album.rs— the Album tab.app.rs—AlbumUiStatefield +Tab::Albumvariant + tab-strip selectable + central-panel dispatch.
9 new tests landed: album::tests covers render-empty errors, single-clip pass-through, two-clip equal-power overlap sums to ~1.0, per-clip gain, and signature stability + change-invalidation. tba::tests covers create+clips round-trip, mix_run round-trip + upsert + delete, reopen-round-trips. tba_album::tests covers Album → .tba → Album. Suite 136 passing (up from 127). fmt + clippy --release --all-targets -D warnings clean under the CI-pinned Rust 1.95.0 toolchain.
- Timeline drag editing (drag clips + fade handles on the timeline strip — Crossfade-tab-style).
.wavas a clip source (v0.4.52 takes.tibonly).- Relative source paths (portable albums).
- Per-clip post-correction at album-render time.
- Probe-each-clip's-decoded-duration on Add so the timeline band reflects real length instead of the 5 s placeholder.
Loading a .tib that has no bounced mix used to return a flat error and force the user to leave the Crossfade tab. v0.4.51 turns that into a guided three-way choice:
- ⤓ Bounce as-is (default focus) — render the master mix using each track's current correction chain. Transparent: tracks with no correction stay uncorrected. Best when you've already tuned the project.
- ✓ Apply Suno-Clean to N, then Bounce — seeds the user's default-correction profile (or Suno-Clean from the built-in presets, or the first available profile — same cascade the Mix-tab "Enable all corrections" button uses) on every track currently without a chain, saves the project, then bounces. Persistent: corrections stay applied.
- ⋯ Open project to tune… — switches to the Mix tab with the
.tibas the active project; user dials each track individually, hits the Mix-tab Bounce there, then reloads in Crossfade. - Cancel — dismiss, no project changes.
Failed bounces (e.g. all tracks muted) keep the modal open with the error inline so the user can pick a different path without re-opening the file picker.
app.rs:CrossfadeBounceFlowstate struct (tib_path / is_a / project_name / track_count / n_without_corr / error) +crossfade_bounce_flow: Option<CrossfadeBounceFlow>field onTinyBoothApp. New methodsbounce_tib_for_crossfade(path, seed_suno_clean)(opens foreign.tibin isolation, optionally seeds the seed-correction profile + saves manifest, renders, writesmix_run— current project untouched) andpick_seed_correction_profile()(same cascade asenable_all_corrections: project default → Suno-Clean → first available).ui/crossfade.rs: pre-flightprobe_tib_for_bounce_flow(path) -> TibProbethat opens the.tib, returnsHasMixRunorEmpty { project_name, track_count, n_without_corr }.handle_loadpopulates the flow state instead of erroring whenmix_runis missing.render_bounce_flow_modal(app, ctx)draws theegui::Windowcentered on screen with the three actions; on bounce-success, re-invokeshandle_loadwhich takes the existing-cache path.
fmt + clippy --release --all-targets -D warnings clean under the CI-pinned Rust 1.95.0 toolchain. Suite 127 passing.
Finished .tib projects become composable units. Bounce once → drop the .tib into the Crossfade tab as a stem → stitch multiple finished pieces without re-rendering the source or staging intermediate .wav files.
Single-row mix_run table holds a 16-bit WAV byte stream plus sample_rate, channels, frames, source_signature, created. The source signature is a stable hash over every project field that affects what the master mix renders (each track's current_rev_id + mute + gain + polarity + automation + correction; project master gain & automation & corrections_disabled). Stamped at Bounce time; compared live on the Mix tab to flag the cache as fresh or stale. Schema bump 1 → 2 with in-place migration — older .tib files get the table added on first open with CREATE TABLE IF NOT EXISTS + a single UPDATE meta SET schema_version = 2.
New button in the Mix-tab transport row. Renders the master mix into memory using the existing mixdown DSP, encodes as WAV, writes to the mix_run row. State pip on the button:
- no cache →
⤓ Bounce - cache present + fresh →
⤓ Bounce ✓ - cache present + stale →
⤓ Bounce ⚠ stale
Disabled on folder-backed projects (the cache lives inside .tib — Save → As .tib first; tooltip says so).
Load… dialog filter widens to WAV or TinyBooth stem (.tib). Loading a .tib opens the SQLite container, reads the mix_run blob, decodes it through the same decode_wav_reader_as_stereo helper the .wav path uses — so zoom, click-seek playheads, fade handles, ms counters, preview, and export all work unchanged. If the picked .tib has no bounce yet: clear status <file> has no bounced mix yet — open the project in TinyBooth and click Bounce first.
tib.rs:MixRunHeaderstruct + four CRUD methods (read_mix_run_header,read_mix_run_audiovia incremental BLOB I/O,write_mix_runupsert,delete_mix_run).export.rs: new publicrender_master_mix+render_master_mix_to_wav_bytes+compute_mixrun_signature(serde-derivedMixSighashed viaDefaultHasherto a 64-bit hex). The existingmixdownstays private — the new entry points just stop at different points in the same pipeline.app.rs:bounce_master_mix_to_tib()+mix_run_status()glue.ui/mix.rs: Bounce button intransport_bar(after the bulk-correction strip).ui/crossfade.rs: extension dispatch inhandle_load+ newload_tib_mix_run_as_stereo.
New mix_run_round_trips_and_upsert_replaces test in tib::tests covers write → read → upsert-replaces → delete. Suite 127 passing (was 126). fmt + clippy --release --all-targets -D warnings clean under the CI-pinned Rust 1.95.0 toolchain.
The 8 px playhead hit zone was a usability dead-end — invisible until you happened to be on it. Replaced with whole-lane seek: click or drag anywhere on Track A → A's playhead jumps to pointer time. Same for Track B. The cursor turns to a crosshair when hovering a lane so the affordance is obvious; the active preview stops on seek (DAW convention).
- Removed the v0.4.45 "drag track B's waveform area to translate B-offset" affordance — it now conflicts with click-to-seek. B-offset is still set via the slider and the per-pixel-precise form value above the lanes; users gain a sub-second seek surface and don't lose offset control.
- Removed the dedicated playhead drag rects — the whole lane IS the playhead's hit zone now.
- Fade handles keep their 10 px hit zones and still win over the lane seek because they're allocated after the lane interactions in the same frame (egui's last-allocated-overlapping-rect-wins rule).
- B's lane click maps the pointer's global time back to B-local seconds via
t_global - b_offsetso B's playhead stays in its own coordinate system (matches how ▶ B seeks).
lane_a + lane_b get click_and_drag interactions handled via clicked() || dragged() + interact_pointer_pos() for absolute jump semantics (not drag-delta). Dead PLAYHEAD_HIT_W constant removed. Suite 126 passing. fmt + clippy --release --all-targets -D warnings + release build all clean.
The playheads were positional indicators only — ▶ A / ▶ B / ▶ Crossfade always rewound to 0 and there was no way to audition just the fade transition without listening from the start. Combined with the zoom strip, this made fine fade-tuning a pain. Fixed.
- Each lane's playhead is now its own transport's seek-to point. Drag A's playhead to 12.345 s, press ▶ A → playback starts at 12.345 s. Same for B's playhead and ▶ B. For ▶ Crossfade, the topmost (A) playhead is the master cursor — it's translated to the mix output frame so the crossfade transport starts wherever A is. The earlier "rewind to 0 on press" behaviour is gone.
- Implementation:
CrossfadePreviewSession::play(...)takes a newstart_frame: u64and initialises its cpal position atomic there. The existingsync_playheads_from_previewalready divides position by sample rate to get seconds, so nothing else needed to change downstream. - Playheads stay visible when zoomed in tight. Previously the playhead disappeared the moment it fell outside the visible view — invisible to the user. Now: if the head is left of view, a small ◀-pointing arrow lives on the lane's left edge; if it's right of view, a ▶-pointing arrow on the right edge. The arrows tell you which direction to scroll/pan to find the playhead.
- Playhead markers on the zoom strip. Both A's and B's playhead positions render as small triangle tips on the global-timeline strip (A points down from the top edge, B points up from the bottom), so you can always see where each head sits on the full timeline regardless of zoom.
CrossfadePreviewSession::play signature gained start_frame: u64. start_preview_track and start_preview_mix compute the start frame from playhead → sample-rate-scaled. draw_lane always draws the playhead (with off-screen edge arrows); zoom strip gets two new triangle markers. Suite 126 passing. fmt + clippy --release --all-targets -D warnings + release build all clean.
Sub-second crossfade work on minutes-long material was effectively impossible at the v0.4.46 scale — a 60-second timeline rendering in ~800 px gives ≈75 ms per pixel, which means the fade handles can't resolve anything finer than that. The new zoom turns the entire lane area into a sub-section of the global timeline so 100 ms / 50 ms / 10 ms fades become draggable directly.
- Zoom strip (16 px minimap drawn above the lanes). It always shows the FULL timeline with thin per-track shadow lines so you can see where you are. Drag any horizontal range on the strip and the lanes zoom to that range. The current view is outlined on the strip in cyan so you can re-orient at a glance.
- Form fields, always visible:
Zoom start(seconds, DragValue) +Zoom %(0.1 … 100, DragValue). Both editable whether you're zoomed in or not — type a value to enter zoom precisely, or scrub. - ✖ Reset zoom button — only enabled while zoomed. One click returns to the full view.
- Nothing else changes its behaviour. Sliders, fade-handle drag, playhead drag, ms counters, ▶ A / ▶ B / ▶ Crossfade preview, Export — all of it works exactly as before, just remapped to the zoomed pixel scale. Drag a fade handle while zoomed in 100× and you get 100× finer per-pixel resolution.
Centralised in a view_range() helper called once per frame; everything downstream consumes view_start / view_dur instead of tl_start / tl_dur. Fade-handle hit zones constrained to the lanes area so they don't bleed into the strip's drag rect. Strip drag uses egui drag_stopped() (renamed from the deprecated drag_released()).
zoom_start_secs, zoom_pct, zoom_drag_anchor_secs added to CrossfadeUiState (defaults: 0.0 / 100.0 / None — no behavioural change for users who don't touch the new controls). Suite 126 passing. fmt + clippy --release --all-targets -D warnings + release build all clean.
Each lane now carries a draggable playhead and a mm:ss.mmm counter so the operator can see and seek to any point in either track from the timeline alone — no transport-relative math, no audio-burn-just-to-find-the-spot. Layered on top of the v0.4.45 UX pass.
- Vertical playhead per lane with a small triangle cap and a narrow 8 px hit zone (sits inside its lane; never conflicts with the existing 10 px fade-handle hit zones, and takes priority over the lane-B translation drag).
- Drag a playhead to seek that track. Dragging stops any active preview — matches DAW seek convention.
- Auto-follows the active preview. ▶ A drives A's playhead; ▶ B drives B's; ▶ Crossfade drives both (each computed as global time minus that track's start, clamped to its own [0, duration]). Preview position is pulled from the cpal session's
Arc<AtomicU64>frame counter via a newsync_playheads_from_previewhelper, and the UI requests a 16 ms repaint while audio is flowing so the heads move smoothly. mm:ss.mmmcounter drawn top-right of each lane in monospace so the digits don't jitter as they tick over.- ▶ A and ▶ B both rewind their lane's playhead to 0 on press, so each transport press is unambiguously start-to-end.
CrossfadePreviewMode enum + preview_mode, a_playhead_secs, b_playhead_secs added to CrossfadeUiState. draw_lane() extended with playhead rendering; new fmt_ms helper formats the counter. Layering of hit-rects keeps fade handles' priority intact. Suite 126 passing. fmt + clippy --release --all-targets -D warnings + release build all clean.
The Crossfade tab gets a proper direct-manipulation surface in place of the slider-only MVP, and the underlying mix model switches from "fade = overlap" to a flexible transition model.
- Drag track B's waveform anywhere on its lane to set the start offset. The slider stays as a fine-tune backup; the two stay in sync.
- Drag fade-region handles independently. Two vertical handles (orange grip caps) mark the fade's start and end. Drag either to size the fade where you want it — the fade no longer has to span the entire overlap. A 30 s overlap with a 1 s fade in the middle is now expressible; so is an instant cut (
fade_start == fade_end). - Transition model: before
fade_startonly A plays (B muted even if present); afterfade_endonly B plays (A muted). Inside the fade range, both contribute via the equal-power or linear curve. This is the standard DAW behaviour — what you'd want for using the tab as a transition tool, not a layering tool. - Snap fade to overlap button resets the fade range to span the current A/B overlap (the previous MVP default, now an explicit affordance).
- Loading a track auto-snaps the fade range to whatever overlap then exists, so the default still does the obvious thing.
- Status row shows the current fade as
fade_start → fade_end (duration).
Three consecutive releases (v0.4.42 / 0.4.43 / 0.4.44) the ship script's poller died mid-poll with no error output, while the actual CI build + publish succeeded fine. Wrapped the gh release view invocation in an outer try/catch so any per-poll transient (deprecation warning, API blip, anything that surfaces through $ErrorActionPreference = 'Stop') gets caught and the loop retries on the next poll. The deadline still bounds total wait. No effect on the release itself; eliminates the false-failure notifications.
DSP rewritten around fade_start_frame_abs + fade_end_frame_abs independent of track positions (CrossfadeSpec and CrossfadeMix updated). 5 unit tests rewritten / added to cover the transition model: equal-power weights sum-to-1-in-power, linear amplitude sum, silent → silent, fade outside overlap mutes non-active track correctly, zero-length fade is an instant cut. UI gains drag-detection via ui.interact(rect, …, click_and_drag) for track B's lane and the two handle rects, with cursor-icon hints. Suite 126 passing. fmt + clippy --release --all-targets -D warnings + release build all clean.
A dedicated tab for two-track crossfades. Load any two WAVs from disk, position track B's start relative to A with a slider, audition the result, and export the mix in any of the formats export.rs already supports (WAV / FLAC / MP3 / Ogg / Opus / M4A).
- Sources — independent of the active project. Two "Load…" buttons, file dialog → decode-once → cache as stereo f32. Mono inputs are duplicated to L=R on load. Sample-rate mismatch between A and B is refused with a clear status message (the player has no resampler — same constraint as everything else).
- Timeline — single shared time axis showing both waveforms with the overlap region highlighted and the fade curve drawn faintly over the overlap. The axis spans
[min(0, b_offset), max(a_duration, b_offset + b_duration)]so negative offsets (B before A) work too. - Curve picker — Equal-power (
cos²/sin², default — sums to 1 in power, right for unrelated material) or Linear (sums to 1 in amplitude, right for phase-coherent material). - Transport —
▶ A,▶ B,▶ Crossfade,■ Stop. Each▶press builds a freshCrossfadePreviewSessionover its own cpal output stream (matched to the source rate; falls back to device default). Drop = stop. The preview is auto-released as soon as playback reaches the end of its buffer. - Export — format combo +
Export…button. Renders the full timeline at the curve and writes via the sharedexport::write_crossfadehelper (same WAV / ffmpeg pipeline the Export tab uses). Default filename<A-stem>_x_<B-stem>.<ext>.
- Drag the track-B waveform left/right to set the offset directly (slider gets us shipping).
- Resampling so mixed-rate WAVs can load.
- Two-stage transitions (fade-in at the start of A + fade-out at end of B on top of the A↔B crossfade — that's a full timeline editor).
New module src/crossfade.rs (pure DSP), new src/crossfade_player.rs (minimal cpal output session), new src/ui/crossfade.rs (the tab). 5 new DSP tests (no-overlap concatenation, equal-power weights sum to 1 in power, negative-offset timeline shift, silent-input → silent-output, linear-curve amplitude sum). Suite 126 passing. fmt + clippy --release --all-targets -D warnings + release build all clean.
Two feature requests' worth of work batched into one release: the Record-tab recordings browser (TBSS-FR-0008 items 1/2/4 + 3-partial) and the Generator-track MVP (TBSS-FR-0009 all 6 steps).
- Loose WAVs (not in manifest) — every
*.wavin the recordingstracks/directory that isn't tracked in the manifest is now listed in its own group with filename / size / mtime / 📂 reveal-in-Explorer. Files dropped manually or carried from another machine are no longer invisible..swap-tmpdebris is filtered out. - Path-label affordances — the "Each take saves to " line grows a 📋 copy-to-clipboard button and a 📂 open-in-Explorer button (auto-creates the folder on first run).
- Waveform thumbnails per take — every row in the recordings list shows a 140 × 28 green-on-dark envelope thumbnail, peaks computed on first visit and cached on the app. Zero-sample takes render as an empty black rect — a perfect at-a-glance signal of orphan / partial-write artefacts.
- Click-drag region selection — drag on a thumbnail to pick a
[start, end]range (translucent orange overlay + thin edge lines). Right-click clears. - 💾 Export Selection — per-row button (disabled until you pick a range) that lossless-crops the selected region via
trim::crop_wav_bytes(16/24/32-bit int + float supported) and writes it to a user-picked path. The original take is untouched.
audio::start_recordingno longer leaves a header-only~44 Borphan WAV intracks/when the cpal stream build or.play()fails after the WAV writer was created. Any failure now finalises the writer +fs::remove_files the partial file before propagating the error. The underlying cpal repeat-take race still needs a real-hardware repro to fix at root — when it next happens you'll see the actual error string in the status bar instead of a silent orphan.
File → "Add Generator Track…" opens a modal that lets you synthesise a focus-music stem (binaural beats or isochronic tones) and bake it on demand into the project, alongside your other stems. The bake reads its duration from the longest other track, renders the audio via pure DSP, and stores the WAV through the existing audio path (.tib revision or folder WAV) plus a timestamped copy under <project>/exports/generator-bakes/<id>-<ISO8601>.wav for a versioned library of bakes.
- Binaural mode — sine carrier with a slight L/R freq offset (
carrier ± beat/2). Continuous, no clicks. Needs headphones. Carrier 40–800 Hz, beat 0.5–40 Hz, amplitude 0..1. - Isochronic mode — sine carrier × a smoothed
sin²(π·φ/duty)pulse envelope. Works over speakers. Tone 40–800 Hz, pulse 0.5–40 Hz, duty 0.05–0.95, amplitude 0..1. - Layered focus music — the third architectural slot from the RFC is surfaced (disabled) so the data model + UI don't need rework when the layered-pad DSP design lands.
- "Meld with master chain" interpretation — Reading A (confirmed during design): the bake snapshots a
MasterSignaturefor dirty detection only; it does not pre-apply the master chain. Playback routes the generator through master like any other track. If you change master settings after a bake, the track marks itself dirty and the user re-bakes. - Locked-track guards —
Track::is_locked()is true for generator tracks. Trim silently skips them (their bytes are bake output, not WAVs you crop); hot-load swap bails with "generator tracks are baked, not swapped — change parameters and re-bake instead". Delete-track is unchanged: removing a generator is a valid action; the WAV is regenerable on re-bake.
- Per-lane dirty indicator on the Mix tab when a generator's stamped signature drifts from the current project state.
is_generator_dirty(idx)exists and is correct; the render is the missing piece. - Edit-params affordance on locked tracks (currently: delete + re-add to change a generator's mode / params).
- Loose-WAV Adopt into manifest + ▶ play / 🗑 delete parity.
- Layered focus-music DSP.
New module src/generator.rs (pure DSP); new modal src/ui/generator_params.rs; data model in src/project.rs (TrackSource::Generator, GeneratorMode, MasterSignature, Track::is_locked(), compute_master_signature); bake plumbing in src/app.rs (bake_generator_impl, bake_generator, is_generator_dirty, add_generator_track, resolve_generator_modal). Suite 121 passing (up from 95 at v0.4.41): 16 new tests across the generator data model (4), DSP (8), bake plumbing (3), and locked-track skip (1). All gates green at every commit.
The single-file .tib SQLite container is no longer just an export target — it is a full live load/save format running alongside the legacy folder format. A .tib project does everything a folder project does:
- Open / Save —
File → Openaccepts both*.tiband*.tinybooth; saves to a.tibare one SQLite transaction touching only changed pages (no whole-file rewrite), crash-safe via WAL. - Play — the player reads each track's audio from its
current_rev_idrevision BLOB via incremental SQLite I/O, on the audio owner-thread. - Trim (reversible) — a destructive Trim writes a new revision and repoints
current_rev_id, keeping the last five edits plus the immutableorigas a FIFO history. Roll back by repointing — no byte copy. (Folder projects keep the in-place crop.) - Export — the mixer reads BLOBs; a
.tibexport is byte-for-byte identical to the folder export of the same audio. - Hot-load swap — replacing a track's audio commits a new revision, so the pre-swap take stays recoverable.
- Telemetry / Project Health — analysis now works on
.tibprojects via a BLOB→temp-WAV bridge (extract, analyze, clean up), so.tibis a true peer of the folder format. - Import lands in
.tib— importing a Suno bundle (folder or zip) migrates the imported project to a sibling.tiband opens it; the folder staging stays on disk as a backup. - Migrate-on-open prompt — opening a legacy
*.tinyboothoffers to convert it to a sibling.tib(additive — the folder is kept as a backup) or open it as a folder.
- Mix-tab lane headers no longer look ragged. The per-track mood pip (the small coloured square) was the last item in the header's chip row, so it landed at a different x on every lane depending on how many instrument / key / AI chips preceded it. It's now pinned to the header's right edge, so the pips form a clean vertical column regardless of the chips before them.
- Recording still targets the folder-based recordings filespace (the
.tib↔recordings bridge is deferred). A stem/revision browser UI and a rollback gesture are the next phase; the storage layer already supports them. - Suite 105 passing; all gates (
fmt,clippy --release --all-targets -D warnings,test, release build) clean.
File → "Export as single .tib…" packs the whole project — every stem plus the bundled Suno mixdown — into one self-contained .tib file (a SQLite database; see TBSS-FR-0007). Each stem is stored as its orig revision, the immutable baseline for the revision history coming in the next phase.
- Additive and non-destructive: the folder project is left completely untouched; this only writes a sibling artifact. The app still loads/saves the folder format — the live-format flip is the next phase, and this proves migration on real projects first.
- Built on the tested storage layer (TBSS-FR-0007 phases 1/2a/2b): a
TibDbSQLite container (WAL,page_size=16384,auto_vacuum=INCREMENTAL, incremental BLOB I/O), the Project↔SQLite mapping, and the folder→.tibmigration.
A survey of how real apps use ZIP for live storage (OOXML/ODF/Krita/Sketch all full-rewrite on save; EPUB/USDZ are read-mostly) plus phase-1 prototyping confirmed ZIP can't do cheap in-place updates or duplicate-name overwrites — fatal for a frequently-saved file with per-stem revisions. SQLite is the canonical application-file-format for this (Audacity 3.0 made the same folder→single-SQLite move). The full rationale is in the RFC.
New deps: rusqlite (bundled SQLite — no system dependency). Storage layer has 11 unit tests (container round-trip, large-BLOB incremental I/O, FIFO-5 prune, Project mapping round-trip, folder migration). The export was verified on the real 9-stem project: it packs into a single 332 MiB .tib and a 36 MiB stem BLOB round-trips byte-for-byte. Suite 95 passing; all gates (fmt, clippy --release --all-targets -D warnings, test, release build) clean.
v0.4.39 moved audio init off the UI thread so it could no longer freeze or crash the app — but the lanes still waited on the output device: the whole Mix display (waveforms, faders, headers) was gated on the player, which wasn't ready until the cpal stream was built. On a flaky/slow output driver you'd stare at a near-empty Mix tab (or the (rendering Mix tab via ctx_panels) placeholder) for the whole device-connect — which feels like a freeze even though the window is technically live.
- The audio build is now two-phase. Phase 1 loads the track WAVs and assembles the shared state — no audio device touched — and hands the UI a player immediately, so the lanes render as soon as the audio decodes, whether or not an output device is present or healthy. Phase 2 probes the device and builds the cpal stream in the background.
- Playback (Play/Stop) is gated on the stream, not on the display. While the device connects you get a "connecting audio output… (mix is ready; playback enables when the device is up)" hint; the rest of the Mix tab — scrolling lanes, faders, A/B, correction editor — is fully interactive throughout.
- No output device? The mix still shows, with a red "no audio output device — Retry" banner instead of a blank tab. This is exactly the requested behaviour: render the mix, banner the device. Retry re-probes the device (handy after plugging in / switching hardware).
- Remaining wait is just the WAV decode (a few seconds for a full multi-stem project); the ~30 s device-enumeration stall no longer blocks the display at all. (Caching peaks in the manifest for sub-second lane render is a possible follow-up.)
All CI/release gates run clean locally before tag (cargo fmt --check, cargo clippy --release --all-targets -- -D warnings, cargo test --release 84 passing, cargo build --release). Verified live: Mix shows "loading tracks…" → lanes render → Play enables once the device connects; the lane render no longer waits on the device.
A bundle release that turns the cross-band-coherence work from v0.4.35–37 into something you act on, plus a serious audio-robustness fix. (Supersedes v0.4.38, whose tagged build never published — it tripped the CI cargo fmt --check gate. The code is identical here, correctly formatted, with the audio + tofu fixes below added on top.)
The Mix tab built its player synchronously on the UI thread (Player::new → cpal output-device enumeration). On a flaky / virtual / mid-negotiation output driver — Bluetooth Hands-Free endpoints and vendor virtual mixers are the usual suspects — that call could hang ~30 s, freezing the whole window and tearing the lane headers mid-paint, or outright panic, unwinding out of the eframe event loop and silently killing the session (a GUI-subsystem build has no console, so it vanished with no trace).
- The build is now asynchronous on a dedicated audio owner-thread. The UI snapshots the project (cheap, no I/O) and hands the slow WAV decode + flaky cpal enumeration + stream creation to that thread, polling the result each frame and showing an "initializing audio output…" spinner meanwhile. The cpal
Stream(which is!Send) lives and dies on the owner-thread; the UI only ever holds theSendArc<PlayerState>handle. The UI never blocks on the audio device again. - A panic inside cpal is now contained (
catch_unwindon the owner-thread) and degrades to the same graceful "no audio device — Retry" banner an error produces, instead of taking down the app. - New panic logger (
main.rs): any panic, on any thread, now appends its message + backtrace to%APPDATA%\TinyBooth Sound Studio\logs\panic.log. A GUI-subsystem build has no stderr, so before this a panic left no trace — this is how the freeze/crash was finally diagnosed.
The amber applied-state badge (below) used ✓ (U+2713), which egui's default font doesn't cover — so it rendered as a □ tofu box, the exact failure v0.4.36 eliminated for the 🤖 emoji. Replaced with the ASCII text "AI on" (basic-Latin glyphs always render).
The Mix-tab "AI" pill is now a button, not just a verdict.
- Click the pink
AIpill on any flagged stem → one-click apply Coherence Restoration to that track. Seeds a correction chain from theSuno-Cleanpreset if the track has none, forcescoherence_restoration.enabled, gives it a sensible default strength (0.5, mid of the recommended 0.3–0.6 range), marks the project dirty, and pushes the snapshot to the player so the next playback cycle hears it. - Once applied, the pill becomes an amber
AI onbadge — so you can see at a glance which flagged stems are already being fixed. Click that to open the correction editor and tune the strength. The badge's tooltip is explicit that the score shown is the raw source measurement and won't change (telemetry runs on the source, restoration runs at playback/export). - Implemented as a deferred action (
ChipActionreturned fromtelemetry_chips, handled after the lane loop) so the immutableplayerborrow held while drawing the lanes drops first — the same pattern theCorbutton already uses.
The 🌀 Visualizer now overlays a live cross-band coherence readout in the top-right of the canvas — the same AI-audio fingerprint metric the analyzer computes per-track, but estimated continuously as you listen.
- Each frame it bins the master-bus spectrum into 6 log-spaced bands (edges 150 / 400 / 1k / 2.5k / 6k Hz → Nyquist), keeps a ~3 s rolling history of per-band energy, and reports the mean of the 15 pairwise Pearson correlations — EMA-smoothed so it doesn't jitter. Cheap: ~1.6k multiply-adds per frame at 30 fps.
- Tiers and colours reuse the shared
telemetry::COH_*thresholds, so the HUD's verdict matches the Mix-tab pill and Project-Health column exactly. Toggle it off with "Live coherence HUD" in the visualizer config panel. - Works across every mode (Lissajous / Mandala / Lorenz / Chladni / Onion Skin) — it's a global overlay, not a new mode.
- The AI / clean tier boundaries (
< 0.45AI,≥ 0.65natural,≤ 0.05not-analysed) were duplicated as magic numbers in two UI sites. They're now named constants —COH_AI_MAX,COH_CLEAN_MIN,COH_PRESENT_MIN— intelemetry.rs, the single source of truth referenced by the Mix-tab pill, the Project-Health column, and the new visualizer HUD. - The constants carry the calibration rationale in-source: anchored to the documented natural (0.6–0.9) and AI (0.2–0.5) ranges and to the DSP test-signal evidence (AI-shaped fixture ~0.33 raw → ~0.52 after restoration; correlated control > 0.8). A restored stem landing in the ambiguous
[0.45, 0.65)middle is by design — restoration nudges it out of the AI band without faking a natural score. Honest caveat noted: values stay heuristic until calibrated against a labelled real-vs-Suno corpus.
The console-strip track-name labels (rotated 90°, reading top-to-bottom) were clipped by a fixed 14-character cap left over from when the label was horizontal — so "Electric Guitar" (15 chars) showed as "Electric Guit…" despite the rotated gutter having ~190–260 px of room. draw_rotated_label now fits the name to the actual rail height, ellipsising only on genuine overflow (operating on chars() so multi-byte names never split a code point). Removed the now-dead STRIP_NAME_CHARS constant and ellipsize helper.
lto = "thin"(waslto = true) in the release profile. Most of the cross-crate inlining win of fat LTO at a fraction of the link time; runtime perf is within noise for this UI + block-DSP workload, and CI link time drops noticeably.
Full suite still 84 passing. All CI/release gates run clean locally before tag: cargo fmt --check, cargo clippy --release --all-targets -- -D warnings, cargo test --release, cargo build --release. The async audio path was verified live — spinner shown during init, lanes render, playhead advances under playback (owner-thread stream confirmed running), no panic.
The payoff of the cross-band coherence arc: v0.4.35 shipped the diagnostic (measures the AI-audio fingerprint), v0.4.37 ships the fix — a real DSP filter that re-correlates the bands.
- New
CoherenceRestorationconfig on everyProfile(enabled: bool,strength: f32).#[serde(default)]→ existing manifests andprofiles.jsonfiles are byte-unaffected; the field defaults to disabled. - The DSP (
CoherenceRestorerindsp.rs): splits the signal into 6 octave-ish bands via a Linkwitz-Riley 4th-order crossover bank (crossovers at 120 / 350 / 1k / 2.8k / 7k Hz), follows each band's amplitude envelope plus the broadband envelope, and applies a per-band time-varying gain that pulls every band's modulation toward the shared broadband modulation shape — while preserving each band's long-term average level, so the spectral balance is untouched and only the modulation correlation changes.strength∈ [0,1] blends the correction gain toward unity;enabled = falseis a true bypass (the crossover bank isn't even constructed). - Chain position: runs after the HPF / DC-remove cleanup, before the EQ — so the rest of the correction chain shapes an already-restored signal. Wired into both
FilterChain(mono) andFilterChainStereo(two independent per-channel restorers, so a hard-panned element on L isn't modulated by R's content). - UI: a "Coherence Restoration" section in the Correction window (and the Admin profile editor) — Enabled checkbox + Strength slider, with hover-text and an inline explanation linking it to the v0.4.35 "Band Coh." telemetry chip.
- Allocation-free hot path: the per-sample
processuses a stack[f32; 6]scratch array; all state (20 biquads + envelope followers per channel) is preallocated at chain construction.
6 new tests in src/dsp.rs, all passing. The headline one: a synthetic realistic AI-shaped signal (shared base modulation + per-band jitter — measures ~0.33 cross-band coherence, squarely in the real-Suno 0.2–0.5 range) run through the restorer at full strength comes out at ~0.52 — a +0.19 gain that clears the 0.45 AI-fingerprint threshold. Also tested: doesn't tank an already-coherent signal, helps even a pathologically-decorrelated one, strength scales monotonically, silence stays silent (no NaN from the envelope-division paths), and build_coherence gates correctly on enabled + non-trivial strength. The proof tests run the restorer end-to-end and re-measure with the actual v0.4.35 compute_cross_band_coherence analyzer. Total suite: 78 → 84 passing.
The restorer's envelope followers (3 ms attack / 30 ms release) and gain smoothing (6 ms) are tuned to track the modulation tightly without audible band-pumping. Pushing harder would raise the measured coherence further but risks artefacts on real audio — the design deliberately stops where the effect is real but clean. Threshold calibration against actual Suno content is the natural follow-up.
- AI-fingerprint chip no longer renders as a
□tofu square. egui's default font has selective Unicode coverage — musical-instrument emojis (🎸 🥁) and arrows (↗) render, but🤖(U+1F916) doesn't and silently falls back to tofu on Suno-shaped stems. Replaced with a plain text "AI" tag rendered in a pink chip (monospace, dark-magenta background, light-magenta foreground) — same colour scheme, same hover tooltip, guaranteed to render in any font. Same change in the Project Health panel's "Band Coh." column. Bundling a Noto-Emoji subset is the longer-term fix but stays out of scope for a small UI patch.
The first half of the cross-band coherence work from docs/sound-vision-philosophy.md §V lands here as a measurement + UI surfaces.
- New telemetry field
cross_band_coherence: f32on every analyzedTrack. Algorithm: pick 8 octave-spaced centres (60 / 120 / 240 / 480 / 960 / 1920 / 3840 / 7680 Hz), sum FFT bin magnitudes in a 1/3-octave window around each centre per STFT frame, z-score the resulting 8 energy envelopes, EMA-smooth (α=0.2 → ~10 Hz cutoff at typical STFT hop), compute pairwise Pearson correlation across all 28 band pairs, return the mean. Cheap — reuses the STFT that was already computed for centroid/onsets/etc.; the extra work is ~8 × N_frames adds + 28 correlations. - What it measures: natural recordings score 0.6–0.9 because every band is driven by the same physical event (one string vibrating, one vocal cord opening — the bands share a common low-frequency modulation envelope). AI-generated audio scores 0.2–0.5 because each band is generated semi-independently and wobbles out of phase. Below ~0.45 is the working threshold for the AI fingerprint.
- Mix-tab chip: a
🤖badge in pink on every track whose coherence < 0.45, with a tooltip explaining the score and reserving phase-3 (Coherence Restoration filter). A≈badge in cool green when ≥ 0.65 — surfaces the "this stem is naturally recorded" case for tracks where it matters (e.g. a real vocal you tracked yourself, dropped into a TinyDAW project alongside Suno stems). - Project Health column "Band Coh." with the numeric value coloured by tier (pink < 0.45, gray middle, green ≥ 0.65) and a
🤖/≈glyph appended at the extremes. - Schema bumped 3 → 4 so existing v0.4.34 manifests are auto-stale and re-analyze on next open. Migration is invisible.
- 3 new tests in
src/telemetry.rs: synthetic common-envelope STFT → coherence ≥ 0.7 (natural case), decorrelated-envelope STFT → coherence < 0.4 (AI case), degenerate inputs (empty / too-few-frames / all-zero) don't NaN. Suite: 75 → 78 passing.
The Coherence Restoration post-processing filter (re-correlates the bands by gating their modulation envelopes against a shared reference envelope) is the eventual payoff — it turns this diagnostic from "tells you which stems are AI-shaped" into "fixes the bands-don't-move-together signature, taking the stem closer to a real recording". Not in v0.4.35; will land as a per-stem opt-in inside the existing correction chain when shipped.
- Lane overlap finally fixed for real. The v0.4.33 diagnostic build revealed that the three top-level Mix-tab panels (transport, lanes CentralPanel, console) are positioned correctly — green stripes between Frame::groups confirmed the lanes CentralPanel starts cleanly below the transport. The bleed was inside
lanes_view: the first lane row's header was rendered viaui.allocate_exact_size+ui.child_ui(header_rect, …)+ manualset_clip_rect. On the first iteration of the lanes ScrollArea's for-loop, the parent ui's cursor was stale (Y ≈ top of screen), so the computedheader_rectwas at Y ≈ 0 and the child_ui rendered there — overlapping the transport panel. Subsequent iterations (rows 2-9) had a valid cursor by then and rendered correctly inside the CentralPanel. - The fix: replaced the manual
child_ui+set_clip_rectwithui.allocate_ui_with_layout(size, layout, |ui| …), which uses egui's own internal cursor tracking and doesn't suffer the stale-cursor issue. Same(HEADER_W, LANE_H)sizing constraint, same lane-header content; egui handles the positioning correctly across all iterations. - Diagnostic colors from v0.4.33 removed —
render_transport / render_console / render_lanesare back to clean delegate functions.
Five attempts at fixing the Mix-tab lane overlap haven't worked. This release tints each Mix region with a faint coloured overlay (red for mix_transport_panel, blue for mix_console_panel, green for the lanes CentralPanel) and prints the panel's actual y=[top..bottom] h=height rect in the top-left corner of each region. The next screenshot from this build will definitively show whether:
- the panels' real rects overlap (egui's space allocation is wrong), or
- the panels' rects are distinct but the lane content draws outside its rect (egui's clip enforcement is wrong), or
- the lane content draws correctly inside green but the green region itself is mispositioned.
No functional changes; this is a one-release-only diagnostic. v0.4.34 will revert the colors based on what we learn.
- The Mix-tab lane overlap is finally fixed for real this time. v0.4.31 moved the Mix-tab panels to ctx level but bundled all three (
mix_transport_paneltop,mix_console_panelbottom, lanesCentralPanel) into a singlemix::ctx_panels(app, ctx)call placed AFTER the globalbottom_barpanel declaration. That violated egui's strict panel-order requirement: allTopBottomPanel::topcalls must precede allTopBottomPanel::bottomcalls, which must precede theCentralPanel. Withbottom_bardeclared beforemix_transport_panel, the bottom-of-screen space was claimed before the Mix-tab top was, scrambling egui's space accounting — visible as the Vocals lane content rendering at the same Y as the transport bar. - The fix interleaves Mix-tab panel declarations across
app.rs::update()in the correct order:top_bar(menu) — line ~1321mix_transport_panel(Mix-tab top) — immediately after, whenmix_activebottom_bar(status) — line ~1591mix_console_panel(Mix-tab bottom) — immediately after, whenmix_activeCentralPanel(tab body) — always last; hosts lane stack for Mix or the tab body for everything else
mix.rsnow exposes the helper API the new layout needs:pre_render(app),compute_console_h(app, ctx),render_transport(app, ui),render_console(app, ui),render_lanes(app, ui). The previousctx_panels(app, ctx)wrapper is gone — it was the wrong abstraction (couldn't be placed correctly relative to the global panels).
- Lane content rendering at wrong Y coords — root cause finally found and fixed. The bug was visible across v0.4.29 / v0.4.30: the first lane's name + chips + M/S/A/B/+Cor row would render at the top of the screen overlapping the global menu bar, while the
Frame::groupborder for that lane was either missing or in its proper place below the transport bar. Different parts of the same row drawing in different y-coords pointed at a painter-layer / Z-order issue, not just bad rect maths. - The actual cause: egui's painter uses multiple layers (Foreground for widgets, Background for panel fills, Tooltip for popups, plus ScrollArea-internal sublayers). Nesting
TopBottomPanel::show_inside/child_ui+set_clip_rectinside the app's globalCentralPanel::show(ctx, ...)only constrains the immediate layer —ComboBoxpopups,ScrollAreaviewports, and tooltip rendering bypass the child's clip_rect and use the OUTERui's. So the lane'sFrame::group(drawn directly in the immediate layer) sat where it should, while the ComboBox / ScrollArea content for that row ended up unbounded. - Fix: the Mix tab now declares its three panels (
mix_transport_panel,mix_console_panel, lanesCentralPanel) at ctx level, as siblings of the app's global menu bar (top_bar) and status bar (bottom_bar), rather than nested inside the globalCentralPanel. This is the egui-blessed pattern for a multi-pane workspace — egui's panel system composites these cleanly because they all draw to the same level, with no painter-layer mismatch. Newmix::ctx_panels(app, ctx)function called directly fromapp.rswhenTab::Mixis active. app.rsgains a branch: for Mix tab with tracks (and not the Visualizer takeover), it callsmix::ctx_panels(self, ctx); everything else continues to render inside the globalCentralPanelviamix::show(self, ui)(now a thin placeholder for the empty-project case).- Removed
TRANSPORT_BAR_Hconstant and therender_clippedhelper from v0.4.30 — both were ceremony for the failed child_ui approach.
- Lane content no longer bleeds above the transport bar. v0.4.29's nested-panel approach (
TopBottomPanel::top+TopBottomPanel::bottom+CentralPanel, all viashow_inside) misbehaved when hosted inside the app's outerCentralPanel::show(ctx, ...)— the first lane row would render above the lanes region, overlapping the global menu bar. Replaced with explicitchild_uiregions whoseclip_rectandmax_sizeare both pinned to a pre-computedRecttaken fromui.max_rect(). The result: each of the three regions (transport, lanes, console) is a hard-clipped rectangular zone, content overflow is physically impossible, and the layout no longer depends on egui's nested-panel internals being well-behaved. - New helper
render_clipped(parent, rect, id, |ui| …)centralises the child-ui pattern so all three regions share the same clipping contract — no copy-paste of thechild_ui_with_id_source+set_clip_rect+set_max_sizetriple. - Added
TRANSPORT_BAR_H = 56.0so the transport region has a known fixed height, preventing the lanes/console below it from shifting when transport content changes width (e.g. the error banner appearing/disappearing).
Full architectural rewrite of mix::show(). The Mix tab now uses egui's native panel layout instead of stacked allocate_ui_with_layout calls, fixing three independent bugs that all traced back to that approach:
┌─ TopBottomPanel::top ────── transport bar + error banner
├─ CentralPanel ────── lanes (the ONLY surface that takes vertical scroll)
└─ TopBottomPanel::bottom ── spectrum + strip cards (horizontal scroll only)
- Each egui panel owns its own clip rect, so content can no longer bleed between surfaces. The track headers / waveforms physically cannot reach into the transport bar above or the console deck below — that's been the actual bug under "headers bleeding top and bottom" since v0.4.21.
- Each panel owns its own scroll-event hit-testing. Pre-v0.4.29 both the lanes and the console shared the caller's
ui, so a wheel event in the wrong place could shift either. Manifested as "the cards jitter in place when I scroll". - The bottom panel claims an
exact_height(console_h)based onmix_console_fraction; the CentralPanel takes whatever's left. The split no longer floats withui.available_height()(which depended on whatever else the parent had drawn — including the top-bar readings changing width when a digit ticked), so the layout doesn't wobble by a px each frame.
- Lanes —
ScrollArea::vertical().auto_shrink([false; 2]), fills the CentralPanel. - Console deck —
ScrollArea::new([true, false])(explicit hscroll-only + vscroll-disabled) +auto_shrink([false; 2]). Vertical wheel events inside the deck no longer try to scroll a 0-height extent, which was the actual cause of the "jittering in place" visual.
show()extracted into four cohesive helpers —rebuild_player_if_needed,render_player_error_banner_if_present,consume_autoplay_request, plus the three panel renderers. No more 130-line function with player-lifecycle, autoplay, layout maths, and scroll plumbing tangled together.- Manual drag handle for the lane↔console split removed. The split is now fully driven by
mix_console_fraction+CONSOLE_H_MAX. If interactive resize comes back, it'll be via egui's nativeTopBottomPanel::resizable(true)rather than a hand-rolledallocate_exact_size+drag_delta().
- Player lazy-rebuild + error banner + Retry button.
- Auto-play hand-off from Record-tab ▶ clicks.
- Automation arm / capture / commit loop on per-strip and master.
- Spectrum panel position (top of the bottom panel, pinned above the strip cards).
- Telemetry chips, profile dropdown, M/S/A/B/Cor row, hot-load swap, every per-track button — unchanged.
- 75 tests still passing; clippy clean with
-D warnings.
cargo clippy --release+cargo test --releasedropped fromrelease.yml. Both gates already run on every PR viaci.ymlin debug mode (faster, catches the same correctness issues). Re-running them in release mode at tag-push time was redundant paranoia — by the time a commit hitsmainand gets tagged, it has already cleared CI. Saves ~3–5 min per release.cargo-wixinstall cached by~/.cargo/bin/cargo-wix.exe, keyed on the toolchain version. v0.4.27 and earlier compiledcargo-wixfrom source on every build. Saves ~1–2 min.- FFmpeg LGPL download pinned + cached. Pre-v0.4.28 the workflow pulled
https://.../releases/download/latest/...every build — 120 MB over the network for the same binary. Now pinned toautobuild-2026-05-12-13-59(a dated BtbN tag, stable) and cached by that key. The asset URL is resolved from the GitHub API rather than guessed, so the file-naming churn between autobuild tags doesn't break the workflow. Bump the pin intentionally when a new FFmpeg version is wanted. Saves ~30–60 s.
Net: ~12 min builds → ~5–7 min builds, zero binary changes. The release artifact is byte-equivalent to v0.4.27 modulo the version-string bump.
- Admin → Audio devices… modal. Picks both the master input device (used by the Record tab) and the master output device (used by Mix-tab playback) from cpal-enumerated lists. Each pick shows the device name plus its native channel count and sample rate. Empty pick = "follow the platform default" — useful when the user wants to track Windows' active default rather than pinning a specific device.
Config.input_device/Config.output_device(Option<String>, both#[serde(default)]). Both persist toconfig.jsonso the picks survive app restarts. Older configs auto-migrate toNone(= follow platform default), which preserves v0.4.26 behaviour exactly for everyone who hasn't touched the new panel.- Graceful fallback when a saved device disappears. When
Config.input_deviceoroutput_devicereferences a name that no longer matches any enumerated device (user unplugged the USB mic between sessions, switched ports, etc.), the resolver falls through to the platform default rather than erroring out. New helpersaudio::input_device_by_name(Option<&str>)andaudio::output_device_by_name(Option<&str>)centralise this lookup. - Rescan button in the panel — re-enumerates cpal's device list mid-session. Plug in a USB mic, click Rescan, the dropdowns update without an app restart.
Player::newsignature gainsoutput_device_name: Option<&str>. Threads through tobuild_output_stream, which now callsaudio::output_device_by_nameinstead of the hard-wireddefault_host().default_output_device(). Same fast-fail probe at the top ofPlayer::newso a missing chosen device still cheap-errors before the WAV-loading phase.- Mix-tab player rebuild triggered automatically when the output device is changed mid-session — drops
app.playerso the next Mix-tab frame rebuilds with the new device. Playback stops; user hits Play to resume. - Record tab's input-device dropdown now restores its previous selection from
Config.input_deviceat startup, falling back to the platform default if the saved device is no longer enumerated.
- Clicking the version label always does the round trip, even when an update is already known to be available. Pre-v0.4.26 the click handler was gated on
state == Idle— so once an "v0.4.x available — click to install" badge appeared next to the version label, clicking the label itself did nothing. That made the label feel half-broken: it advertised itself as click-to-refresh, but at the very moments you'd want to refresh (e.g. "is there an even newer release than the one shown?") the click was inert. Now the click forces a freshcheck_latest_release()whenever the updater isn't in the middle of a check or a download (those two states are still guarded so we don't race on the receiver). Hover-text added to the label spelling out the behaviour.
- Lanes no longer overlay the transport bar. The Mix tab used
TopBottomPanel::top("mix_lanes_panel").show_inside(ui, ...)for the lanes block andTopBottomPanel::bottom(...)for the console deck.show_insidepositions panels using the parent ui'smax_rect, IGNORING the current cursor — so the lanes panel landed at the absolute top of the central area, painting over whatever thetransport_barhad drawn there. Visible in the v0.4.24 screenshot as the Vocals lane bleeding over theMix · Pause · Stop · Enable corrections …row. Replaced bothTopBottomPanel::show_insidecalls withui.allocate_ui_with_layout(...)— that respects the cursor, so the lanes sit cleanly below the transport bar and the console deck cleanly below the resize handle.
- Strip cards no longer staircase diagonally in the console deck.
ui.horizontaldefaults toAlign::Centeron the cross axis, so cards with even slightly different effective heights drifted down each step (visible as a Vocals → Backing Vocals → Drums → Bass cascade in the v0.4.23 screenshot). Switched toui.with_layout(Layout::left_to_right(Align::Min), ...)— every card's top edge now sits on the same y baseline. - Lane header bleed killed. v0.4.21's 1-px line divider between rows was too subtle to break the eye-fuse between adjacent headers. Each lane is now wrapped in its own
Frame::groupwith a dark fill, so every row is a visibly bounded card with a clear border on all sides. Removed the redundant divider. - First lane no longer kisses the transport bar above it — added 4 px of explicit top padding inside the lanes ScrollArea so the Vocals row's name has breathing room from the controls row above.
scripts/ship.ps1poll fixed. v0.4.23's first run sat polling for 14 minutes on a release that had been published 6 minutes earlier. Two bugs: (a)gh release view --json publishedAt, assetswas being parsed by PowerShell as two arguments, andghrejected the second one (Unknown JSON field: " assets") with a non-zero exit code that the script swallowed via2>$null. Fixed by quoting the field list as a single token ('publishedAt,assets') and gating on$LASTEXITCODE. (b) The published-detection check used a regex^20\d\d-against$pubAt, but PowerShell 7'sConvertFrom-Jsonauto-parses ISO 8601 strings into[DateTime]objects whoseToStringproducesMM/dd/yyyy …in US locales — the regex never matched. Replaced with a non-null check.
- In-app updater no longer goes stale for the session. The long-standing known issue tracked since v0.4.12 — bottom-bar version label showing the install version indefinitely because
check_latest_releasefired only once at app startup — is closed. Newgit_update::maybe_spawn_recheckruns every frame, rate-limited atRECHECK_INTERVAL = 300 s, gated onstate == Idle && rx == None. Two triggers force a non-rate-limited recheck: (a) the 5-minute timer expiring, (b) any tab transition (Record ↔ Project ↔ Mix ↔ Export). The check itself is a single small JSON GET, so the work is bounded; the gate guarantees we don't fire while a previous check is still in flight or while the user is mid-update. ~25 LOC insrc/git_update.rsplus two new fields onTinyBoothApp(last_update_check_at: Option<Instant>,last_tab_seen: Option<Tab>). No new dependencies.
scripts/ship.ps1— PowerShell script that owns the full "tag pushed → MSI downloadable" arc, not just the push. Pushes main + the tag, then blocks pollinggh release view <tag>every 15 s untilpublishedAtbecomes a real ISO timestamp, then prints the asset SHA-256 fingerprints and download URLs. Hard 30-min timeout so a stuck CI run can't hang the script forever. Closes the operator-side half of the same gap the updater fix closes on the app side: before today, "ship" wasgit push --tagsplus a vibe — no signal that the release-build pipeline was healthy and the artifact was actually downloadable. Usage:.\scripts\ship.ps1 -Tag v0.4.24.
- Strip-card name labels rotated 90° (top-to-bottom reading), sitting in a narrow gutter on the left edge of every card instead of a horizontal centred label across the top. Saves a full row of vertical space inside each card (the fader rail claims that height now), and matches the classic mixing-console label-runs-along-the-side aesthetic. Implemented via egui's
epaint::TextShape::angle = π/2. Same treatment on the master strip with its existing yellow accent. - Playback readings collapsed into the top bar as a right-hand aside. Time (pos / dur), sample rate, and momentary / integrated LUFS used to live in the Mix-tab transport bar, taking up a row of horizontal space. They now sit in monospace font on the right side of the top menu bar next to the project name, visible from every tab (was Mix-only). Format:
M ±NN.N I ±NN.N LUFS 48000 Hz 02:06/03:20. Fixed-width per-field padding so the digits don't jitter. - Transport bar slimmed. With readings out, the bar is now a tight strip of controls: Play / Pause / Stop · Enable corrections · Disable (saves) · Reset · A/B. No more 1280-px-wide row of mixed-info-and-controls.
- Strip-card button row tightened: M / S / R / Ø shrunk 22×22 → 18×20 so they fit alongside the new label gutter inside the fixed STRIP_W = 108. Inner margin tightened 8 → 6 for the same reason.
- Strip cards no longer balloon on tall windows. v0.4.19's "stretch fader rail to fill available height" change ran unbounded — on a window where the console-deck pane was 500+ px, each strip's fader rail grew to 400+ px and the cards became "gigantic" (user screenshot). Added two hard caps:
FADER_H_MAX = 200(rail max) andCONSOLE_H_MAX = 340(whole deck max). The drag handle between lanes and deck still adjustsmix_console_fraction; the caps just prevent the deck from eating the screen. Net effect: tall windows get more vertical space for the lane stack (which is where you actually mix), less for the strips (which don't need 400 px of rail). - Lane headers stop bleeding into each other. v0.4.18's
LANE_H = 52was a hair too tight for the 2-row header (name + chips above, M / S / A/B / Cor + profile dropdown below), so adjacent rows visually fused with no clear boundary. BumpedLANE_H52 → 62 andROW_GAP4 → 8, and added a 1-px horizontal divider centred in the row gap. Each lane is now a clearly bounded card with comfortable padding.
- Hot-load: ↔ Swap audio. New button on every Project-tab track row. Pick a WAV; the bytes replace the track's audio in-place, preserving every other field on the manifest — track name, role, correction chain, volume automation, polarity flip, telemetry profile. Sample-rate enforcement: the new file has to match the project's existing rate (TBSS still has no resampler; mismatched rates would break the Mix tab silently). On mismatch the swap is refused with a clear status, nothing on disk changes. On success the project is auto-saved, the player drops itself so the next Mix-tab frame rebuilds with the new audio, telemetry is invalidated and re-dispatched, and the project-level Krumhansl-Schmuckler key estimate is recomputed because old pitch histograms no longer apply.
- Transparent TinyBooth metadata injection — every hot-loaded WAV gets a TBSS JSON blob written into its standard RIFF
LIST/INFO/ICMT(comment) field before the file goes live. The blob carries: project name, source classification (Suno role / Recorded / TinyDAW take), polarity-inversion flag, active correction-profile name, telemetry profile, and atinybooth-sound-studio v0.4.20produced-by string. Any RIFF-aware reader (exiftool, foobar2000, our ownsuno_meta::read_wav_session) sees a standard comment; TBSS sees a structured record it can round-trip. New modulesrc/wav_meta.rswithinject_tbss_meta(write side) +read_tbss_meta(read side, reserved for the upcoming "drop a WAV onto the Project tab → mint a track preserving TBSS context" feature). - Atomic on-disk writes. Hot-load swap and metadata injection both write to a
.swap-tmp/.tbss-tmpsibling then rename over the live file, so a process crash mid-swap can never leave a half-written WAV in the project folder.
- File → New TinyDAW project… — creates a non-Suno, recording-centric project. The Mix tab, Export, Health panel, telemetry, automation, correction-chain UX are all identical; what changes is the routing rule: a Suno / untitled project sends captured takes to the canonical recordings filespace at
%APPDATA%\TinyBooth Sound Studio\recordings\, while a TinyDAW project receives its takes directly into its own folder. Switches to the Record tab on creation so the next click ⏺ goes into the new project. - New
Project.kindfield —ProjectKind { Standard, Recordings, TinyDAW }.Standard(default) preserves v0.4.19 behaviour for every existing manifest. The canonical Recordings filespace gets tagged asRecordingson its next open (one-time migration via the existingopen_or_create_recordingspath; the field is#[serde(default)]so older manifests don't reset).
- Both hot-load swap and transparent TBSS metadata injection apply identically to TinyDAW projects — they're per-track operations that don't care about Suno context. A TinyDAW user can swap any recorded take with a different WAV (e.g. replace take 3 with the cleaner take 5, keeping its correction chain), and every WAV the project produces carries the project provenance in its RIFF comment.
- 3 new unit tests in
src/wav_meta.rs: inject+read round-trip (with hound reopening the file to prove the WAV is still valid), repeated injection replaces the previous blob (no unbounded file growth),read_tbss_metareturnsNoneon a plain WAV. Total suite: 72 → 75 passing.
- Spectrum panel relocated from top-of-Mix-tab to top-of-console-deck. Sits directly above the fader strips now, so the meter ↔ spectrum comparison happens in one glance instead of being a screenful apart. The console deck's vertical budget grew as a result — see strip redesign below. Toggle remains under Admin → Show spectrum panel (Mix tab).
- Strip cards now stretch the fader rail into their full vertical space. Pre-v0.4.19 the fader was pinned at
FADER_H = 130regardless of how tall the console-deck region was, leaving a wide blank zone below the dB readout on tall windows. The strip (and the master strip) now computefader_h = available_h − 110so the rail + peak meter fill whatever's left after the label / button-row / dB-readout claim their fixed share. Floors at the oldFADER_H = 130so a too-short console deck still shows a usable fader. The peak meter scales with it so a louder signal now sweeps the full height of the card, matching the recalibrated spectrum panel.
- Spectrum panel was completely saturated all the time. Root cause in
analysis::spectrum: the FFT bin magnitude wasn't window-corrected. For a 0 dBFS sine at FFT bin centre, the raw Hann-windowed bin reads asN/4 ≈ 1024(atN = 4096), and20·log10(1024) ≈ +60 dB. The old((db + 80) / 80)mapping clamped that to1.0immediately, so any real music content pinned every bar at the top. Two-line fix: multiply the magnitude by4 / N(Hann amplitude-coherent-gain inverse) so a 0 dBFS sine actually reads as 0 dBFS, then map((db + 90) / 100)for the bars —-90 dB → 0,0 dB → 0.9(10% headroom at the top for transient overshoots). Existing spectrum-floor and peak-bin-position tests still pass. - Strip-card bottom space no longer wasted. Same issue as above — fader was fixed-height inside a taller frame. Fixed by the fader-stretch change.
- Mix-tab spectrum panel — pinned at the top of the Mix tab when enabled (default on). Live FFT of the master output bus drawn as bars on a log-frequency X axis (20 Hz → 20 kHz, with 100 Hz / 1 kHz / 10 kHz decade gridlines), normalised log-mag Y axis, plus a slow-release peak-decay trail (0.95×/frame ≈ 1 s release at 30 fps) sitting above the live spectrum. No new audio-thread plumbing — reads the same
PlayerState.output_vizmaster-bus tap that v0.4.11 added for the standalone visualizer canvas. New module src/ui/spectrum_panel.rs. - Admin → Show spectrum panel (Mix tab) checkbox toggles the panel on / off, persisted via
Config.show_spectrum_panel(defaulttrue,#[serde(default)]so oldconfig.jsonfiles don't reset). v0.4.18 adds the field; older installs gain it on first save.
- Mix-tab lane headers compacted from 3 rows → 2. Pre-v0.4.18 each lane header was three rows tall (name + profile dropdown / chips / M·S·A/B·+Correction), needing
LANE_H = 72. New layout: row 1 = track name + telemetry chips, row 2 = M / S / A/B / Cor + profile dropdown.LANE_Hdropped 72 → 52,ROW_GAP6 → 4. Net: ~28% more lanes visible per screen height. The "+ Correction" button label shortened to "Cor" / "+Cor" so the row fits inside the existing 240-pxHEADER_Wwithout crowding the dropdown. Hover-text on the button still carries the full explanation.
- Drum classifier no longer over-counts events ~3-6×. The v0.4.13 multi-band onset detector emitted one event per (band, frame) pair — so a single snare hit, which produces real flux peaks in MID + HIGH_MID + HIGH simultaneously, generated separate Snare + Cymbal + HiHat events. Real-world numbers on a 3:20 Suno drum stem: ~5,300 total drum events ≈ 27/sec (physically impossible — sane rate is 3–8/sec).
classify_drum_eventsnow flattens every per-band onset into a single time-sorted candidate list, clusters candidates whose frames fall within 3 of each other (the same< 3window the universalall_onset_framesdedup already uses inanalyze_wav), and per cluster picks the dominant band by normalised flux strength (raw flux / band's flux max — the only fair cross-band comparison since absolute flux magnitudes vary wildly between low and high frequencies). The dominant band's frame alone runs through the existing kick/snare/hat/tom/cymbal classification. Total drum-event count on a typical 3-minute stem now lands in the 500–1500 range, distributed across classes — the same order of magnitude as the (correctly deduplicated) universaltel.onset_count. ANALYZER_VERSIONbumped 2 → 3. Existing v0.4.13–16 manifests are treated as stale and re-analyzed on next project open. Migration is invisible — the dispatcher already skips up-to-date rows. Drum-event chips and Project Health rolls-ups will repopulate with sane counts after the first re-analysis pass.
- New unit test
drum_classifier_dedupes_per_hit_no_double_count: synthesises a 1-second WAV with one kick (60 Hz sine, 50 ms decay, sub-band only) at 0.2 s and one snare (broadband xorshift noise burst, 40 ms decay, fires MID + HIGH_MID + HIGH simultaneously) at 0.6 s. Asserts the drum classifier produces exactly 2 events total — the snare cluster collapses correctly. Pre-v3 this test would have failed with 4–6 events. - Total suite: 69 → 70 passing.
- Per-channel
M(mute) andS(solo) buttons on every Mix-tab lane header. Previously only available in the console deck strips at the bottom of the tab — invisible while working in the lane view. Now mirrored at the lane level next toA/Band+ Correction. The atomic flags (track.mute,track.solo) are shared with the console-deck strip + the audio thread, so flipping in one place reflects everywhere immediately.
- Lane waveforms now share a common X-start across every row. v0.4.15's
allocate_ui_with_layout(vec2(HEADER_W, LANE_H), …)was a suggested size — when the inner content's natural width exceeded HEADER_W (chip strip, profile dropdown text), it grew the box and pushed the lane allocation right by a handful of pixels. Every row's waveform / playhead landed at a slightly different X. Replaced withallocate_exact_size(…)+ achild_uiwhose clip-rect is set to the header rect, so any inner overflow is hard-clipped and the lane allocation begins at exactlyHEADER_Wpast the row's start regardless of telemetry density. HEADER_Wbumped 220 → 240 andLANE_Hbumped 60 → 72 to give the new third row of buttons (M / S / A/B / +Correction) breathing room without crowding the profile dropdown above.
- Mix-tab telemetry chips: single-line consolidated strip. v0.4.13–14 rendered each telemetry feature as its own chip — bright/dark + sustained/percussive + dense + 5 separate drum-class counts (
K744 S1789 h1288 T1280 C232) + guitar pick chip + bend chip + key chip + mood pip. On drum/percussion stems the chip strip wrapped onto a second line, making the row taller than non-drum rows — every lane started at a different vertical offset, layout looked uneven. Replaced with a fixed three-element strip rendered viaui.horizontal(no_wrapped, no overflow): one instrument summary chip (🥁 5333or🎸 730 ↗3) carrying the full per-class breakdown in its tooltip, one key chip when confident, one mood pip whose tooltip carries every spectral / dynamics / rhythm numeric that used to be a separate chip. Every row is now the same height regardless of telemetry density. - Tooltips on every header-column control — telemetry profile dropdown now has a hover explaining all six profile options (Auto / Universal only / Drums / Guitar / Bass / Off) plus the currently-active resolved profile.
+ Correction/Correctionbutton has hover explaining what attaching a chain does (and what happens when the seed comes from the project default vs. the Suno-Clean preset). Mood pip's tooltip got expanded from one line to a structured block with mood / timbre / dynamics / rhythm sections so glance-decoding the colour is feasible.
Flagged as a follow-up: the multi-band onset detector emits independent events per band, so a single snare hit produces concurrent events in MID + HIGH_MID + HIGH and gets counted as Snare + Cymbal + HiHat. Real numbers on a 3:20 Suno drum stem: ~5,300 total drum events (≈ 27/sec, physically impossible — sane rate is 3–8/sec). The universal onset_count is correctly deduplicated; the drum classifier needs the same cross-band time-window dedup with class arbitration via dominant flux peak. Targeted for v0.4.16. Doesn't affect tonality / non-drum analyses.
- Guitar / bass pick-stroke detection with YIN pitch tracking (TBSS-FR-0005 phase 2). Each spectral-flux onset in the MID + HIGH_MID bands becomes a candidate event; YIN runs on a 50–150 ms post-onset window for sub-sample-accurate pitch; a polyphony probe counts spectral peaks above –12 dB to flag strums. Each event is classified into one of:
Pluck— single-string monophonic pick at a new pitchRepeat— same pitch as previous (within ±50 cents, configurable) — tremolo / repeat pickingStrum— polyphonic onset; no pitch reported, single event per strum (per the design discussion's "1 event per strum" decision)Slide— smooth pitch trajectory continuing from the previous event, between 50–200 cents in <100 msNoise— onset detected but velocity below the configured pick threshold OR YIN gave up cleanly
- Pitch persisted as raw Hz plus YIN confidence (cmnd at the chosen lag). Cents-off-pitch / detune analysis / bend density / key inference / riff fingerprinting are all free post-processing of the persisted data — no re-analysis needed when those features land.
- Krumhansl-Schmuckler key detection (per-track + project-level). Per-track: weighted pitch-class histogram (velocity × duration-until-next-pitched-event) → 24-key Pearson correlation against the canonical Krumhansl & Kessler 1982 templates → top key + runner-up. Project-level: union of every guitar/bass track's histogram, recomputed every time a guitar/bass result lands in
drain_telemetry_results. Surfaced on the Project tab ("Estimated key: G♯ min") and per-track on the Mix-tab lane chips ("♪ E♭ maj"). - User-selectable analyzer profile per track (TBSS-FR-0005 §"Phase 2"). New
▾ Auto/▾ Guitar/▾ Bass/▾ Drums/▾ Universal only/▾ Offdropdown on every Mix-tab lane header. Default isAuto— resolves from the track'sStemRole(drums → drum kit, electric/acoustic guitar → guitar, bass → bass, everything else → universal-only). Explicit values override — useful when Suno mislabels a stem asFxOtherwhen it's actually a percussive synth, or when a recorded take has no role at all and the user wants a guitar pitch read on it. Changing the profile clearstrack.telemetry, persists, and re-dispatches. - Admin → Telemetry settings… modal with sliders for every analyzer threshold:
- k·MAD onset threshold (default 3.0)
- Guitar pick velocity threshold (default 0.05)
- Bass pick velocity threshold (default 0.04)
- YIN cumulative-mean-difference threshold (default 0.15)
- Same-pitch tolerance in cents (default 50, controls Pluck / Repeat split)
- Polyphony cutoff (default 5 peaks above –12 dB → Strum)
Persisted to
%APPDATA%\TinyBooth Sound Studio\telemetry_settings.json. Snapshotted into eachTelemetryRequestat dispatch time so in-flight analyses use the values that were active when they were queued (mid-batch edits don't corrupt running work).
- Project Health panel gained two columns:
Profile(selected → resolved, e.g.Auto → guitar) andKey. Instrument-layer column now shows🎸N ↗N (poly NN%)for guitar tracks alongside the existing drum-kit roll-up. ANALYZER_VERSIONbumped to 2 — old v0.4.13 telemetry is treated as stale and re-computed on next project open. Migration is invisible: the dispatcher already skips up-to-date rows.
Suno's bundle is purely audio (WAV stems + RIFF LIST/INFO/ICMT provenance — already parsed in src/suno_meta.rs). No .mid, no symbolic notes, no chord chart. Suno is a generative audio model; stem separation is post-hoc Demucs-style source separation that by construction can't recover MIDI. All pitch data has to come from our own analysis — YIN is the lever. A future pitch_source: Analyzed | ImportedMidi enum is reserved on the schema so user-supplied sidecar .mid files (from Basic Pitch / Melodyne / MT3) can plug in without a re-migration.
- 7 new unit tests:
- YIN recovers pure 440 Hz within 5 cents on a synthetic A4 sine
- Polyphony probe scores chord (313 + 461 + 727 Hz, no clean common period) ≥ 0.1 higher than a pure 440 Hz sine
- Krumhansl-Schmuckler returns root=C, mode=Major on a hand-built C-major-scale histogram (confidence > 0.7)
- K-S returns None on an all-zero histogram (no /0)
KeyEstimate::labelproduces "C maj" / "A min" / "A♭ maj" for canonical rootsTelemetryProfile::resolvehonours explicit values (Guitar over a Drums-roled stem) and Auto resolves correctly- End-to-end: synthetic 3-pitch guitar-like WAV (decaying sines) → Guitar profile → ≥2 picks detected, ≥1 pitched event recovered
- Total suite: 62 → 69 passing
- Per-track audio telemetry — pure-DSP analysis baked at first save (TBSS-FR-0005). Every imported stem and every recorded take is now analyzed in the background by a dedicated worker thread and the result is persisted on
Track.telemetryinside the.tinyboothmanifest. No ML, no LLM, no service calls — just rustfft + a single STFT pass per track. The first phase ships these features:- Spectral character: spectral centroid (brightness), spectral flatness (Wiener entropy — tonal vs. noisy), 85% spectral rolloff. Means and standard deviations across the track.
- Dynamics: RMS dB (mean + stddev), peak dBFS, crest factor (peak / RMS).
- Rhythmic articulation: spectral-flux onset detection with adaptive median + k·MAD threshold. Reports onset count, onset rate (Hz), and a sustain ratio (fraction of frames within 10 dB of the loudest moment).
- Mood proxies: arousal in
[0,1](weighted blend of RMS, onset rate, centroid) and a phase-1 valence stub in[-1,1](centroid × tonality). Surfaced as a small coloured pip whose hue tracks valence (cool blue ↔ warm yellow) and whose saturation tracks arousal.
- Drum-kit class detection for stems whose role is
DrumsorPercussion— gated on role per the design spec, so the kick/snare/hat classifiers never run on vocals or pads. Algorithm is multi-band parallel onset detection (Option B from the design discussion): one STFT pass, five frequency-band energy curves (SUB 40-120Hz,LOW_MID 80-300,MID 200-800,HIGH_MID 1.5k-5k,HIGH 5k-12k), per-band spectral-flux onset detectors fire independently. Each event gets:- class (Kick / Snare / HiHat / Tom / Cymbal / Other) decided by which bands the onset lands in plus a harmonic-content test (HNR > 15 in the 100ms post-onset window) for kick-vs-tom disambiguation,
- velocity normalised flux peak,
- decay_ms measured peak → 30 % energy.
- Mix-tab telemetry chips under each track lane's name, pulled from the manifest. Phase-1 chip vocabulary:
☀bright,🌙dark,≈sustained,⚡percussive,▦dense. Drum stems additionally show counts:K12 S8 h31 T2 C4. Hover any chip for the underlying numerics. Mood pip on the right edge of the chip strip. - Project Health panel (Project tab → "📊 Project Health…"). Modal showing per-track analyzer status, mood readout (arousal · valence), drum-event roll-up, and metadata weight in bytes (computed via JSON serialisation of each
TrackTelemetry). Where "Infinity events vs. cap" got resolved: no event cap, but the user can see the cost and decide whether to compact in a future build. Live "Analyzing N/M…" progress while a batch is in flight, also surfaced as a chip on the bottom-bar. - Background telemetry worker (
crate::telemetry::TelemetryService) — single named OS thread, owns onempsc::Receiver<TelemetryRequest>and onempsc::Sender<TelemetryResult>. UI thread dispatches at every lifecycle event that produces a fresh WAV (Suno import,stop_take, project open, project re-open after Trim) and drains results inupdate(), patching the matching tracks and saving the manifest once per drain. Foreign-project results (e.g. Recordings analysis lands while the user has a Suno project active) get written through to the recordings manifest on disk so nothing is lost. Cost target ≈ 1-3 s per 3-minute mono stem on a modern CPU; runs at idle priority through the OS scheduler since the audio callback never sees this thread. - Schema version on telemetry (
analyzer_version: u32) so future analyzer changes can detect stale rows and re-compute on demand. The dispatcher already gates on this — tracks at the current version are skipped on every "open project" pass. Initial schema is1.
- TBSS-FR-0005 was written before this build (full RFC at
docs/feature-requests/TBSS-FR-0005-track-telemetry.md). The implementation ships phase 1 + drum-kit detection together as decided in the design discussion. Phases 2-4 (pitch tracking, key detection, cross-band coherence, visualizer integration) remain queued.
- 6 unit tests in
src/telemetry.rs: silence handling without panic, pure-tone brightness detection, transient detection (≥3 of 5 synthetic pulses), arousal monotonicity, valence clamps, drum-class glyph non-emptiness. Total suite count up from 56 → 62.
- New visualizer mode: "Onion Skin" (multi-timescale trajectory) — addresses the long-standing critique that audio visualizers are derivative-of-NOW and never show the volumes / cadences / colors a listener navigates while listening. Plots
(spectral_centroid, RMS)as motion through 2D feature space with three layers of temporal memory: bright recent trail (note / beat scale), faded ghost trail (phrase scale, ~30 s), and a session-wide residency watermark heatmap (section / song scale). Optional anticipated-future projection extends the trajectory linearly via recent-direction averaging. The first mode designed against the "memoryless visualisation is sterile" critique articulated indocs/sound-vision-philosophy.md. Axis labels (soft↔loud, dark↔bright) so the listener orients at a glance. - Collapsible left-side config panel in the visualizer screen exposing every per-mode parameter as a slider / checkbox. Every control has a
.on_hover_text(...)helper explaining what it does, what good values look like, and where defaults came from. Toggle visibility via the new "◀ Hide config" / "▶ Show config" button next to the heading. - Temporal smoothing for the modes that benefit:
- Mandala — exponential moving average on the spectrum (default α=0.6, slider 0..0.95). Reduces jitter without losing responsiveness; reveals the steady-state structure underneath the transient flicker.
- Onion Skin — EMA on the (centroid, RMS) point before plotting (default 0.5). Trades note-level reactivity for trajectory readability.
- Per-mode parameter structs (
LissajousParams,MandalaParams,LorenzParams,ChladniParams,OnionSkinParams) onVisualizerParams. Defaults reproduce v0.4.11 behaviour exactly; existing users see the same modes unless they tweak the new sliders. - New top-bar "Hide config" / "Show config" toggle and per-mode hover descriptions on the mode buttons.
- New essay:
docs/sound-vision-philosophy.md— long-form engagement with the question of what it means to transform sound into vision. Argues most audio viz is sterile because it operates only at the sample / note timescale while listeners parse music at five hierarchical timescales simultaneously. Maps "volumes / cadences / colors" onto those timescales. Develops the "onion skin" insight (each visualised moment contextualised by its neighbors across multiple timescales). Includes a substantial DSP detour: the v0.4.11 Mandala's visible jerkiness on AI-generated audio is a real diagnostic signal — AI audio has band-decorrelated micro-fluctuations where natural recordings have correlated ones. Sketches a "Coherence Restoration" filter as a v0.5+ feature that would smooth this signature in the modulation domain, taking AI output meaningfully closer to "sounds like a recording". Linked from the README's contributor docs.
- README's contributor-docs section now links the new philosophy essay alongside
design-vibes.md.
- 🌀 Audio-reactive visualizer — toggleable full-window canvas accessible from the menu bar. Click the icon to take over the central panel; click again to return. Four mathematically-grounded modes, all on egui's 2D painter (no GPU shaders, no extra deps):
- Lissajous goniometer — XY plot of master-bus L vs R with phosphor-trail alpha gradient. Reveals stereo image geometry at a glance: mono content draws a vertical line, anti-phase draws a horizontal 45° line, full stereo draws organic figure-8s. Crosshair guides for the canonical phase angles.
- Spectral mandala — radial FFT, frequencies arranged around the centre with magnitude as petal length. Mirrored across the X axis for mandala symmetry. Hue tracks frequency: warm reds at the bass end, cool cyans at the treble. Tonal balance becomes literally glanceable.
- Lorenz attractor (audio-modulated) — RK4 integration of the Lorenz ODE with σ / ρ / β tugged in real time by spectral centroid and RMS. The strange attractor breathes with the music; auto-fitting projection keeps the orbit centred regardless of parameter drift. Trail of 2000 points coloured by recency through a hue gradient.
- Chladni cymatics — superposition of ten of Ernst Chladni's classic eigenmodes (sin·sin combinations on the unit square) weighted by FFT band energies. Renders the actual mathematical eigenmodes Chladni discovered in 1787. Slow phase drift keeps the figure animated even on steady-state input.
- Master-bus sample tap on
PlayerState.output_viz—parking_lot::Mutex<VecDeque<(f32, f32)>>of lengthOUTPUT_VIZ_LEN(4096 stereo frames, ~85 ms at 48 kHz). Audio thread pushes post-fader L/R samples on every callback; UI thread snapshots when rendering the visualizer. Brief lock window keeps the audio callback within budget. - 4 unit tests on the visualizer's pure helpers: RK4 Lorenz integrator stays bounded over 10k steps on default chaotic parameters; HSV → RGB conversion handles primary colours and zero-value black correctly.
The mode selection synthesises a brief literature scan covering: cymatics / Chladni patterns (CymaVis, Cymatica), audio-reactive Lorenz visualizers (3D Music Visualizer, Cherry Audio's Lorenz module), radial-FFT analyzers (audioMotion, WaveForge), and phase-space portraits in audio research (Audio Visualization in Phase Space, Royal Society's phasegram paper). Reaction-diffusion (Gray-Scott) was considered and skipped — needs GPU shaders to run real-time at fullscreen, scope creep for v0.4.x.
- Bundled static-LGPL ffmpeg. TinyBooth's MSI now ships a
ffmpeg.exenext to the main binary, sourced from BtbN/FFmpeg-Builds's nightly LGPL build. FLAC / MP3 / Ogg Vorbis / Ogg Opus / M4A-AAC export Just Works on a fresh install — no separate download, no PATH plumbing, no scavenging binaries off the internet. Trade-off: install size jumps from ~10 MB to ~130 MB. CI workflow downloads + extractsffmpeg.exetotarget/release/beforecargo wixpackages the MSI; newbinary_ffmpeg<Component>inwix/main.wxsreferences it. License attribution lives in the README's "Built with" line and the Export-tab manual chapter — TinyBooth uses ffmpeg as a separate subprocess (the LGPL-compliant integration mode for non-free apps). - Update-download dialog with rotating fortune-cookie tips. The bigger MSI means a longer self-update download; the existing tiny
"downloading…"label in the bottom bar got old fast. Newsrc/ui/update_dialog.rsshows a centred modal overlay duringUpdateState::Downloading(_)with a spinner, a one-line note explaining why the download is heftier, and a rotating tip card cycling through 22 workflow facts every 6 seconds (recordings filespace, polarity flip, LUFS targets, F1, per-role presets, the cleanse, trim, automation arm, A/B, coherence, Suno-X chains, recordings list ▶, DC remove, polarity-as-debug-tool, etc.). Hooked fromapp::update()after the bottom-bar render — no-op whenupdate_stateisn'tDownloading.
- README's "What it does" Export bullet, manual chapter
01-getting-started.md, manual chapter06-export.md, andappendix-a-troubleshooting.mdrewritten to reflect the bundledffmpeg.exe. The manual now distinguishes MSI-installed copies (ffmpeg is there, do nothing) from source builds (legacy fallback paths still apply). - README's "Built with" line gains the FFmpeg attribution + LGPL pointer.
- Missing audio output device froze the Mix tab and spun fans. When
default_output_device()returnedNone(no headphones, sound card disabled, etc.), the failure happened at the end ofPlayer::new— butPlayer::newhad already loaded every track WAV into memory by then (~600 MB ofi16sample arrays for a typical 9-stem Suno project). On Err, those allocations got dropped. The Mix-tab lazy-rebuild then re-calledPlayer::newon the next frame becauseapp.player.is_none(). Result: 600 MB of WAV decode + allocation per frame, allocator pegged, UI frozen, fans on full. Two fixes:Player::newnow probes the output device first, before any WAV loading. Fast-fail on no device — bails in microseconds with a clear message ("connect headphones or speakers (or check Windows sound settings) and click Retry above") instead of allocating half a gig only to throw it away.- Per-frame retry storm killed by failure cache. New
app.player_attempt_failed_for: Option<PathBuf>records the project root that the lastPlayer::newattempt failed on. The Mix-tab rebuild guard checks against the current project root and short-circuits when they match. Auto-invalidates on project change (root path comparison). Manually invalidated by a new↻ Retrybutton rendered next to the error banner — the natural recovery path when the user plugs in headphones.
- The Mix-tab error banner gains a
↻ Retrybutton when there's a failed-rebuild cache. Click rebuilds the player; the natural path back to a working Mix after fixing audio hardware externally.
- Cleanse hoisted from Mix-tab to top of
app::update(). Previously the cleanse only ran insidemix.rs::show(), so a user who landed on the Project tab first saw their orphans untouched until they specifically clicked Mix. The cleanse is now a once-per-frame call at the top ofupdate()regardless of active tab — orphans clear out the moment a project is open. Cheap-path cost is oneiter().any()overproject.tracks(microseconds with the v0.4.7 perf fix already in place); no observable cost on idle Project / Export / Record tabs. - Missing-source orphans are now dropped from the manifest cleanly. When a
Recordedorphan's WAV file no longer exists on disk (user moved it via Explorer, manual delete, etc.), the cleanse used to: tryrename→ fail ENOENT, trycopy→ fail ENOENT, push a "could not move" failure into the report, and restore the orphan intoproject.tracks. Result: a forever-failing cleanse, status-bar full of red errors, manifest stuck pointing at a ghost. v0.4.8 detects missing source upfront viasrc_abs.exists(); the orphan gets dropped from the manifest with no migration attempt and the count goes into a newremoved_missing_countfield onCleanseReport. Status surfaces as"Cleanse: removed 1 dangling manifest entry/entries (source WAV missing)". Clean, terminal, no retry loop.
- Mix-tab CPU / fan-spin. Three per-frame allocation hot paths killed perf on the Mix tab — measurable as fans spinning up after a few seconds on the tab:
lanes_viewcalledtrack.correction().is_some()andtrack.automation().as_ref()once per track per frame. Both methods take aparking_lot::Mutexlock and clone the entire contents —Profile(Strings, 4-band EQ array, de-ess fields) andAutomationLane(Vec<AutomationPoint>). With 12 tracks at 30 fps that's 720 Profile clones + 720 AutomationLane clones per second, all heap allocation.TrackPlaynow exposeshas_correction(&self) -> bool(atomic-bool mirror, no lock) andwith_automation<R>(&self, f: impl FnOnce(Option<&AutomationLane>) -> R) -> R(callback-style borrow, no clone). Lanes view switched to both.Profile/AutomationLanecloning is now zero per frame on the Mix tab's idle path.cleanup::cleanse_recordings_in_suno_projectran on every Mix-tab frame and unconditionallydrain()-ed + rebuiltproject.trackseven when no orphans were present — pointless heap shuffling on the common-case path. New cheap pre-check (tracks.iter().any(|t| matches!(t.source, Recorded))) returns the empty report before any mutation when there's nothing to do.
- The
correction()/automation()methods are kept onTrackPlayas#[allow(dead_code)]for non-hot-path callers (project-save sync, future diff logic) — clone-via-lock is the right shape for those, just not for per-frame UI peeks.
- MSI installer relaunches the app on successful install. The in-app self-updater spawns
msiexec /passiveand exits so the install can replace the running .exe — but the MSI then ended silently, leaving the user staring at an empty desktop with their session gone. v0.4.6 adds a Type-18 custom action keyed off the installed exe that runs at the end ofInstallFinalize, so the new version comes up automatically and the user lands back where they were. Gated onUILevel >= 3 AND NOT Installed— fires on/passive(the self-updater path),/qr,/qf, and standard double-click installs; skips/qnsilent corporate deploys, repairs, uninstalls, and modify-installs. Runs under user-context impersonation so the app comes up at the user's integrity level, not elevated.Return="asyncNoWait"so msiexec doesn't sit blocked waiting for the app to close.
Player::new's per-track conformance check now covers BOTH rate AND length. Previously a track was skipped only on rate mismatch (and on file-load failure). Suno stems are co-rendered, so they share a single rate and a single length — a stem whose length differs from the rest by more than 100 ms is by definition an alien (orphan recording, different-generation take, etc.) and gets the same skip-and-warn treatment as a rate mismatch. The first successful track sets the project's reference rate + reference length; subsequent tracks must match within tolerance. Status-bar warning surfaces both reasons when both fail:"skipped track 'X': rate Y Hz vs project Z Hz; length F1s vs project F2s". Tolerance was chosen to absorb codec-level packet-alignment jitter that legitimate Suno output may exhibit (sub-millisecond) without letting through actual orphans (typically seconds different).
Player::newis now tolerant of per-track failures. Previously, one missing or unreadable WAV (or a single rate-mismatched row that the cleanse couldn't reach) aborted the whole player and the Mix tab dead-ended on a red error banner. Now each load failure is sent through the audio-error channel as a "skipped track 'X' (file): " warning that the status bar surfaces, and the player builds from whatever tracks loaded successfully. The fail-fast Err is reserved for the case where no track loaded at all.- Mix tab no longer early-returns on
player_error. A partial player still renders its console; the error banner stays as a warning above the transport bar instead of replacing it. Combined with the tolerantPlayer::new, you can mix the surviving stems while seeing exactly which row went bad. - Full anyhow error chain in the player-error banner.
format!("{e}")only printed the top-level wrapper ("reading track …/track-010.wav") with the actual hound failure (file missing? corrupt header? path mangled?) hidden in the chain. Switched toformat!("{e:#}")so the underlying cause renders inline. - Mix-tab rebuild loop on permanently-broken track rows. The lazy-rebuild check compared
state.tracks.len()(post-tolerant-load survivors) againstproject.tracks.len()(manifest count). With one track permanently failing to load, those values never matched and the player rebuilt every frame — re-loading every WAV every render, re-sending all warnings every render. NewPlayer.project_track_countfield captures the manifest count at build time; the rebuild check keys on that, so the broken-track case stabilises after one rebuild.
- v0.4.2's cleanse protocol gated on
suno_mixdown_path: Some(_)to identify Suno projects, but that field only exists on bundles imported in v0.4.0+. Suno projects imported in v0.3.x havesuno_mixdown_path: None(serde default for older manifests) — and those are exactly the projects most likely to contain pre-v0.4.0-bug recording orphans. The cleanse silently no-op'd on every v0.3.x-vintage project. Detection signal expanded: a project is now considered Suno-shaped if EITHERsuno_mixdown_path: Some(_)OR any track carriesTrackSource::SunoStem { .. }. New regression test covers the v0.3.x scenario explicitly.
- Cleanse protocol for legacy bug residue. Pre-v0.4.0, recordings could be appended to whatever project the user had open at capture time — including imported Suno stem projects. The result: a Suno project's
tracks/ended up withTrackSource::Recordedorphans at the wrong rate, breakingPlayer::new's uniform-rate check on the next Mix-tab visit. v0.4.2's cleanse runs at the top of every Mix-tab render: scans the active project forRecordedentries whilesuno_mixdown_path: Some(_), moves each WAV out into the recordings filespace via atomic rename (cross-device fallback to copy+delete), mints freshtrack-NNNids in the recordings project so we never collide with existing recordings, and removes the orphans from the active project. EveryTrackfield is preserved (gain, mute, automation, correction chain, polarity, etc.) — no work lost. Idempotent and cheap when there's nothing to do. - New module
src/cleanup.rswithcleanse_recordings_in_suno_project(&mut Project) -> Result<CleanseReport>and 4 unit tests covering empty-report behaviour, rate-mismatch flag rendering, failure-line rendering, and the non-Suno-project no-op path. - Status bar surfaces a multi-line
CleanseReport.summary()after migration: how many moved, any per-file failures with file name + reason, and a ⚠ warning when migrated takes don't match the recordings project's existing rate (which would break Mix on Recordings).
stop_takenow keepsapp.projectin sync when the recordings project happens to be the active one (via File → Open Recordings). v0.4.0'sstop_takesaved the new take to the recordings manifest on disk but never updated the in-memoryapp.project, so a user who had the recordings project open and recorded a take saw "the take disappeared" until they reopened the project. Mirrors the existing pattern indelete_recording. Player is also dropped so it rebuilds with the new track count on the next Mix-tab visit.
A focused minor release built around the bundle → cleanup → mix → release path. Eleven per-role correction presets, import-time coherence verification, polarity flip, DC trim, Nyquist cleanup, BS.1770 LUFS metering, project-trim panel, and a dedicated recordings filespace with paged Record-tab list. Reference playback A/B and the multi-take browser are deferred to v0.5.0.
- Recordings now live in a dedicated, app-owned filespace at
%APPDATA%\TinyBooth Sound Studio\recordings\, hosting a single persistent.tinyboothproject that accumulates takes across sessions. Captures never contaminate the active stem-mixing project (Suno bundle or otherwise) — recordings and stem mixing are separate concerns. NewConfig::recordings_root()helper andProject::open_or_create_recordings()constructor. - Record tab redesigned: the existing recorder header (profile / device / source / name / ⏺-⏹ / live waveform / spectrum / level meters) sits at the top, and a new paged "Recordings" list takes the rest of the tab. Each entry shows name (hover for the on-disk path), duration, mode, and the recording-tone profile; ▶ button sends a take to the main mixer in one click (swaps
app.projectto the recordings project, switches to Mix, solos that take, starts playback); 🗑 deletes the WAV + manifest entry. Pagination at 10 entries per page, newest first. - Mix-tab autoplay hand-off — new
mix_autoplay_pending+mix_autoplay_solo_idxfields onTinyBoothApp, consumed by the Mix-tab show() right after the lazy player rebuild. Solos the target track, rewinds to position 0, and starts playback in one go. Single-frame transition; the user clicks ▶ on a recording entry and hears it through the same console as their stem mixes. - Recording-rate enforcement:
audio::start_recordinggains arequired_sample_rate: Option<u32>parameter. The Record tab now keys this on the recordings project's existing rate (the rate of the first take ever captured into it), so subsequent takes always match. Cpal refuses up-front rather than landing a broken WAV and breaking the Mix tab on the recordings project later. Project'stracksis loaded fresh on everystart_new_take/stop_takeso the manifest stays the single source of truth — no in-memory dual-project state to drift. File → Open Recordingsmenu entry swapsapp.projectto the recordings project. Same shape as the existingOpen project…flow but skips the recents-list bookkeeping (recordings are scratch, not user-curated).
- New isolated trim panel opened from the Project tab via a
✂ Trim project…button. Single batch operation: pick a[start_secs, end_secs]range, hit Apply, and every WAV in the project (stems + the bundled Suno mixdown) is cropped in place atomically (.tmpsibling + rename so a crash mid-write leaves either the old or the new file intact). Coherence analysis stays valid post-trim because every file in the project shares the same new frame-0. - Concept and waveform-rendering pattern adapted from the sibling
SoundTrimmerproject; integration is intentionally lightweight — no per-track offsets, no manifest-schema changes, no engine surgery. The trim panel is modal-style and doesn't weave into the Mix tab. Per-track trim offsets and drag-handle visual selection are deferred to v0.5.0. mm:ss.mmmtime entry with live parse feedback and over-end clamping. Small reference waveform thumbnail behind the start / end markers (drawn from the mixdown if present, else the first track). Failure breakdown in the status row when any file fails to trim, so the user can see which file (and why) without digging in the import log.- New module
src/trim.rs(backend) +src/ui/trim.rs(panel). 6 new unit tests on themm:ss.mmmparse / format round-trip, including the bare-seconds andss.mmm-only formats and the negative / garbage rejection paths.
- LUFS metering on the master bus (BS.1770-4). New
src/lufs.rsmodule implementing the K-weighting filter cascade (pre-filter shelf + RLB high-pass), 100 ms-slice mean-square accumulation, and gated integrated loudness with the spec's −70 LUFS absolute gate + −10 LU relative gate. Audio thread feeds the master bus into the meter once per frame; UI reads the published readouts via atomics. New labelled monospace block on the Mix-tab transport bar: "M ±X.X · I ±X.X LUFS" — momentary 400 ms window plus gated integrated whole-programme. Hover tooltip names the streaming targets (Spotify −14, Apple Music −16, broadcast −23). Reads—until 400 ms have played; resets on Stop. - Mixdown loudness measured at import. New
Project.suno_mixdown_lufs: Option<f32>populated by a one-shotcompute_wav_integrated_lufspass over the bundled mixdown at import time. Logged in the import log alongside the coherence block; lays the groundwork for the matched-loudness reference A/B button (phase 3b). - 5 new unit tests on the LUFS meter: silence integrates to NaN; a 1 kHz tone at −20 dBFS reads near −20 LUFS (within 1.5 LU); +6 dB amplitude shift produces +6 LU readout (verifies the dB↔LUFS arithmetic); momentary / integrated both return NaN before 400 ms of audio.
- Import-time coherence analysis. Every Suno bundle whose extracted contents include a mixdown WAV (filename containing
master,mix, orfinal— the existingStemRole::Masterheuristic) now triggers a coherence pass: sum all stems at unity gain, subtract the mixdown, compute residual RMS relative to mixdown RMS. Below ~−30 dB ⇒ stems compose cleanly; above ~−10 dB ⇒ a stem is missing, mislabelled, length-mismatched, or polarity-flipped. - Per-stem polarity-vs-mixdown check. Pearson correlation between each stem and the mixdown over its active region. Stems with
r < −0.3get flagged with an⚠ ANTI-PHASEbadge in the import log and a "try the Ø button" pointer in the import-result modal. Doesn't auto-flip — that's a user decision — but surfaces the suggestion at exactly the moment the user is reviewing what just imported. - Mixdown stored as project reference, not summed track. The bundled Suno mixdown WAV no longer becomes a regular
Track(which would double the audio when the user hits Play). It's kept on disk in the project'stracks/folder but referenced via a newProject.suno_mixdown_path: Option<String>. The matched-loudness reference A/B button that uses this — switching the bus output between user-mix and bundled mixdown — is deferred to v0.5.0; v0.4.0 ships with the meter (phase 3a) and the import-time mixdown LUFS reading. - New module
src/coherence.rs— streaming f32-mono RMS / Pearson-correlation analysis at a 4 kHz decimation rate (memory bounded regardless of song length). 6 unit tests covering RMS edge cases, identity / inverted / orthogonal correlation, and the verdict-categorisation summary.
- Per-role Suno-X preset library. 11 new built-in presets (
Suno-Vocal,Suno-BackingVocal,Suno-Drums,Suno-Bass,Suno-ElectricGuitar,Suno-AcousticGuitar,Suno-Keys,Suno-Synth,Suno-Pads,Suno-Percussion,Suno-FxOther) with chains tuned for each role's typical Suno artefacts. Added auto-seeding at import: each detected stem gets the matching Suno-X preset as itscorrectionchain on import, so projects open with usable defaults instead of a flat unprocessed mix. Strings/Brass map to the closest existing chain (Pads / Synth respectively); Master and Unknown intentionally stay unseeded. - Two new processing primitives on every
Profile:dc_remove_enabled(sub-audible 5 Hz HPF that strips DC drift AI generators sometimes leave in stems) andnyquist_clean_enabled+nyquist_clean_hz(top-octave LPF, default 18 kHz, that suppresses Suno's characteristic shimmer in the top octave). UI rows for both in the Profile editor (Admin window + per-track Correction window). Signal flow: input gain → DC remove → HPF → EQ → de-esser → gate → comp → makeup → Nyquist clean. Both default off; the Suno-X presets opt in. - Polarity flip per track (
Øbutton on the Mix-tab channel strip; standard audio-gear glyph for phase invert). Persists viaTrack.polarity_inverted: bool. Implemented zero-cost in the player: the per-buffer cache folds the ±1.0 sign factor into the pre-computed static linear gain, and the automation gain branch picks up the same factor — no extra multiplies in the per-frame hot path. - Profile-library forward-migration.
dsp::load_or_seednow appends any built-in preset whose name isn't already on disk, instead of only seeding a fresh file. Existing user-tuned profiles are preserved verbatim; the new Suno-X library is added once, ever, on next launch.
- Mix tab fader sliders rendered as 14-px stubs at the top of their 130-px bounding boxes. v0.3.10 set
ui.style_mut().spacing.slider_width = 14.0thinking that knob controlled rail thickness, but for a vertical slider in eguislider_widthis the main-axis (rail) length — so the rail was clamped to 14 px. Set it toFADER_H(130) so the rail fills the bounding boxadd_sizedallocates. Rail thickness comes from the cross-axis allocation (rect.width() / 4in egui's slider rendering), which is already substantial at the widerSTRIP_Wv0.3.10 introduced.
- View → UI scale slider (0.75×–2.5×, 5% steps, percentage-formatted) so the entire interface — fonts and widget metrics — grows proportionally for high-DPI / accessibility / small-laptop scenarios. Persists via
Config.zoom, applied through egui'sset_zoom_factorso spacing and button hit-targets scale alongside text rather than text-on-tiny-buttons. Reset-to-100% button next to it. .github/workflows/ci.yml— runs the same three quality gates (cargo fmt --check,cargo clippy --release --all-targets -- -D warnings,cargo test --release) on every PR tomainand every push tomain, with concurrency-cancel and doc-only path filtering. Closes the gap that let v0.3.6→.7 and v0.3.8→.9 burn version numbers on toolchain-shape problems a PR-time gate would have caught.
- Mix tab — channel-strip visual pass.
STRIP_W78 → 108 px; track-name font drops.small()for an explicit13.0pt; dB readout 12.0pt monospace; master strip name 14.0pt. M/S/R buttons grow from 20×18 → 26×22 and the row isvertical_centered-wrapped so it sits squarely under the name instead of left-leaning. Slider rail/thumb thickness bumped from the egui ~8 px default to 14 px (scoped per-strip, doesn't leak elsewhere). Frameinner_margin6 → 8 px. Net effect: track names like "Backing Vocals" / "Electric Guitar" / "Synth / Lead" no longer chop mid-word; the dB readout stops wrapping into one-character-per-line stacks; faders read at a glance. - Track-name truncation switched from a 9-byte hard slice (
&name[..9]) to a UTF-8-safe ellipsis helper (ellipsize(name, 14)). The byte slice would have panicked on multi-byte chars like accented vowels or emoji in track names; the helper operates onchars().
Config.zoomnow carries#[serde(default = "default_zoom")]. Without it, anyconfig.jsonwritten before the field existed failed to parse, and the silent.unwrap_or_default()reset every preference (dark mode, recent projects, last project, profile name) on first launch with the new schema. Standard schema-migration discipline; should have been there from day one.
docs/architecture.md §6.2rewritten to cover both workflows and a new §6.2.1 on the sync-tax trade-off (why duplicated gates beat reusable-workflow indirection at this scale, and what to keep aligned acrossci.yml↔release.yml).- Cross-reference comments at the top of
ci.ymland on the toolchain step ofrelease.ymlso drift is visible at edit time.
- CI install regression: pinning
dtolnay/rust-toolchain@1.95.0(v0.3.7) doesn't shiprustfmt/clippyby default — versioned tags require an explicitcomponents:block. v0.3.8's CI failed atcargo fmt --checkwith'cargo-fmt.exe' is not installed. Same content as v0.3.8 (which never produced an MSI) plus a two-line workflow change.
0.3.8 — 2026-04-27 (no MSI; CI failed installing rustfmt)
CHANGELOG.md— this file. Hand-curated; release notes from the GitHub release page remain auto-generated from commit messages.Track::recorded(...)andTrack::from_suno_stem(...)constructors so future schema additions don't fan out to every literal call site.- Profile editor body shared between Admin → Recording-tone profiles… and Mix → Correction… windows via a new
ui::profile_editormodule — single source of truth for the input-gain / HPF / EQ / de-esser / gate / compressor / makeup chain UI.
chrononow ships withdefault-features = false(audit follow-up;clock+serde+stdare the only pieces we use). Smaller dep tree and binary.- CI's Rust toolchain is now pinned (
dtolnay/rust-toolchain@1.95.0) — local-vs-CI clippy drift surfaces at PR time, not at tag-push. Config::savereturnsResult<()>and writes atomically via a.tmpsibling +renameso a crash or full disk mid-write doesn't leave the file truncated. The UI thread surfaces failures via the status bar.export.rs::mixdownno longer pre-multiplies samples by static gain at read time; gain is applied per-frame in the same loop as automation. Drops a ten-line apologetic comment about a "gain-undo trick" the previous shape required.audio.rssample-format dispatch (mono and stereo branches) gains an inline comment explaining why the six near-identical match arms exist: monomorphisation forces one arm per concreteT, and a macro would obscure the call sites for marginal LOC gains. Rated Nit in the audit; this captures the decision in-source.
Track.profileandTrack.correctiondoc comments now explicitly distinguish their roles (recording-time snapshot vs post-processing chain).
0.3.7 — 2026-04-27
- CI clippy regression:
unnecessary_sort_byon Rust 1.95.0 stable. Same content as v0.3.6 (which never got an MSI built — its CI run failed on this lint) plus a one-line.sort_by_key(...)swap.
0.3.6 — 2026-04-27 (no MSI; CI failed on the new gates)
- 27 inline unit tests across
automation,analysis,suno_meta,suno_import,git_update, andproject. Coverage matches the survival guide §9.1 payback list. - CI quality gates:
cargo fmt --check,cargo clippy --release --all-targets -- -D warnings,cargo test --releasebetween version-check and build. - Audio-thread error channel:
cpalerr_fnclosures push through ampsc::Sender<String>; the UI thread surfaces messages in the status bar instead of locking stderr.
git_update::renderreturnsbool(#[must_use]); on a successful installer launch the caller closes viaegui::ViewportCommand::Closeso Drop impls run (WAV writers finalise, config saves). Pre-v0.3.6'sprocess::exit(0)skipped Drop entirely.git_update.rsswitched fromResult<_, String>toanyhow::Result;.map_err(format!)calls become.context(...)chains.- Clippy hygiene: 14 warnings → 0 (redundant closures simplified, manual
div_ceil→.div_ceil(), derivedDefaultimpls, threeelse ifcollapses, four#[allow(too_many_arguments)]on internal helpers). cargo fmtran across the tree; 23 files reflowed.
0.3.5 — 2026-04-27
- "Enable all corrections" button glyph:
+→✓. The plus read as a small cross next to the destructive⟲ Reset; checkmark is the affirmative action.
0.3.4 — 2026-04-27
- Persisted Disable button on the Mix tab. Flips
Project.corrections_disabled, syncsPlayerState.global_bypass. Survives reload — non-destructive project-wide bypass. Project.default_correctionfield. Drives the Enable cascade: existingTrack.correction→Project.default_correction→ feature default (Suno-Clean).
- Existing destructive Disable all button renamed to ⟲ Reset all to clarify it strips chain configs.
enable_all_correctionsnow uses the three-step cascade above.- Phase-B audio-callback refactor: zero per-callback
Vecallocations; per-buffer cache for atomic loads (~250× fewer per typical 256-frame buffer); static fader gain pre-converted to linear once per buffer instead of per-sampledb_to_lin.
0.3.3 — 2026-04-27
- Ephemeral global A/B toggle on the Mix tab transport. Flips player's
global_bypassatomic without touching the project state. Mid-playback, instant.
0.3.2 — 2026-04-27
- Bulk correction toggles on the Mix tab transport:
+ Enable all corrections/− Disable all. Adaptive labels showing how many tracks each affects.
0.3.1 — 2026-04-27
- Suno session metadata captured at import: epoch (Unix integer seconds, sortable directly), ordinal (project-relative monotonic), provenance.
- Duplicate-import detection: re-importing the same Suno render triggers a Replace/Cancel modal before any files are touched.
Project.next_suno_ordinalcounter; bumped on every successful import.
0.3.0 — 2026-04-26
- Console mixer on the Mix tab — vertical fader strips per track plus a master strip with stereo meters, M/S/R toggles.
- Volume automation — fader gestures recorded during armed playback, replayed via Catmull-Rom splines (
splinescrate). Per-track and per-master. Track.gain_automation,Project.master_gain_automation,Project.master_gain_db.
0.2.2 — 2026-04-26
- Suno import was silent on failure. Now lenient (skips bad entries instead of bailing); writes a per-import diagnostic log to
%APPDATA%\TinyBooth Sound Studio\logs\; pops a modal after every import (success or fail) with summary, log path, and Open Log Folder button.
0.2.1 — 2026-04-26
- Auto-restore last project on startup via
config.last_project_path. - File → Open Recent (eight most-recently-opened, dead entries auto-pruned).
0.2.0 — 2026-04-25
- Mix tab with multitrack waveform lanes, synchronized playhead, transport, per-track A/B bypass, Correction editor.
src/player.rs— cpal output stream, pre-loaded track buffers, atomic playhead, transport state.Track.correction: Option<Profile>; mixdown at export honours it.
0.1.6 — 2026-04-25
- DSP substrate from TBSS-FR-0001: parametric EQ + de-esser added to
FilterChain/FilterChainStereo;Suno-Cleanpreset shipped.
0.1.5 — 2026-04-25
- In-app manual: 12 chapters embedded via
include_str!ofdocs/manual/*.md.Help → Manual…orF1anywhere.
0.1.4 — 2026-04-24
- Suno stem bundle ingestion (folder + zip).
TrackSource::SunoStem { role, original_filename }.StemRolecovers the documented 12-stem set plusInstrumental/Master/Unknown.
0.1.3 — 2026-04-19
- Stereo visualisation: dual waveforms, dual peak meters in stereo recording mode.
0.1.2 — 2026-04-19
- Real brand icon (walnut booth + cream mic + teal waveform). Multi-size ICO; window viewport icon embedded in exe; banner README header.
0.1.1 — 2026-04-19
- Stereo recording:
SourceMode { Mixdown, Channel(u16), Stereo }.FilterChainStereowith envelope-linked gate + compressor.
0.1.0 — 2026-04-19
Initial release. Skeleton-bootstrapped Rust + egui app:
- Record tab with cpal input, channel/mixdown selection, recording-tone presets (Guitar default), live waveform + spectrum + peak meter.
- Project tab with track table, JSON manifest format (
.tinybooth). - Export tab: WAV native via hound; FLAC/MP3/Ogg/Opus/M4A via ffmpeg subprocess.
- Self-update via GitHub Releases.
- WiX MSI installer; tag-driven CI.