Measurement-Driven Tuning for Car Audio and Multi-Way Loudspeaker Systems
Automatic time alignment with honest verdicts, a virtual DSP crossover designer, and engineering-grade acoustic analysis — impulse response, frequency response, phase, loopback-referenced timing, and live transfer functions — on Windows.
Measure each driver once, then leave the car: align, combine, and optimize the whole system from your desk — and only then type the result into the DSP.
Download latest release · Your first measurement · Build from source
Resonalyze is an open-source desktop application for measuring and tuning multi-way loudspeaker systems — with a special focus on the hardest room of all: the car cabin. It generates test signals, records the response through a Windows audio device, and turns the captured data into engineering-focused plots and concrete DSP settings: crossover corners, per-driver delays, polarity, and PEQ. The same toolset measures rooms, home loudspeakers, headphones, microphones, and complete signal paths.
Its center of gravity is the step most measurement workflows leave to you: turning a set of per-driver measurements into one coherent system. Auto delay and Auto crossover search the actual settings against the phase-aware predicted sum, and every automatic result comes with an honest verdict — the engine reports why it trusts an arrival, and refuses loudly instead of fabricating a number when the measurement cannot support one.
Resonalyze is under active development. Treat its results as diagnostic measurements, not as certified laboratory data.
Virtual DSP — combine measured drivers through gain, delay, polarity, crossover filters, an all-pass stage, and PEQ before touching the hardware DSP.
Does the automation actually help? Sum loss — how many dB the real phase-aware sum falls short of a phase-blind magnitude addition at each junction (average / worst dip). Left: a three-way system tuned by ear over years. Right: the same system after one Auto crossover + Auto delay pass — the worst dip shrinks from −8.0 to −2.3 dB.
EQ Wizard equalizes any measured response — an impulse response, a captured overlay curve, or a moving-microphone RTA in dB SPL — toward its own target curve. |
Time Alignment estimates loopback-referenced delay from the transfer impulse response, with confidence, levels, distance, and the arrival envelope. |
If you already use REW, OpenSoundMeter, or Smaart: those are broad measurement toolboxes; Resonalyze is a focused, end-to-end tuning workflow for active multi-way systems. REW's alignment tool sums a pair of measurements and its EQ module corrects a response; Resonalyze operates one level up — separate per-driver measurements on one absolute time base, complete virtual DSP chains, a phase-aware sum of the whole system, and optimizers that work every crossover junction and both stereo sides at once. Its home turf is the car, and its output is not just a plot but the DSP settings themselves:
- Built for multi-way active systems — measure each driver separately, then design the whole system virtually: crossover corners, slopes and families, per-driver delay and polarity, all-pass stages, and PEQ, tuned against the phase-aware predicted sum. Auto crossover and Auto delay search these settings automatically, across both stereo sides in one run.
- Honest automation — an automatic tuner that guesses is worse than none. Arrival estimates carry confidence and verdicts, modal build-up latches and playback crosstalk are detected instead of aligned to, and when a measurement cannot support a decision the engine says so.
- Loopback-referenced timing — a recorded loopback channel is the time reference, so delay and transfer-function analysis are tied to the actual playback path, and separate measurements share one absolute time base.
- Repeatable, calibrated measurements — average up to 64 sweeps into one cross-spectrum transfer estimate with a per-frequency coherence (γ²) curve, and put the response in real dB SPL from an acoustic 1 kHz calibrator.
- Crossover summation prediction — the true complex (vector) sum of two measurements accounts for relative delay, polarity and phase the way dB-curve arithmetic cannot, with a companion sum-loss curve; Virtual DSP takes this to its conclusion with complete virtual chains per driver.
- Fast compare-and-adjust work — persistent and calculated overlays, target curves, on-plot labels, and a measurement history that keeps each entry's whole working state.
Resonalyze does not try to be every acoustic tool at once. For room EQ at home, REW remains excellent; when the question is "what delays, crossovers, and polarities do I put into this six-channel DSP", that is what Resonalyze is for.
A one-minute tour of the main features:
Download the latest ready-to-run build from GitHub Releases:
Resonalyze-Setup-vX.Y.Z-win-x64.exe— the recommended installed buildResonalyze-vX.Y.Z-win-x64.zip— for most Windows computersResonalyze-vX.Y.Z-win-arm64.zip— for Windows on ARM
The .zip builds are self-contained and do not require a separate .NET
installation; the installer adds shortcuts, uninstall support, and automatic
in-app updates for the installed x64 build, and a SHA-256 checksum file is
provided with every release. "Self-contained" refers to the runtime, not to your
data: by default every build keeps settings, history, overlays, Virtual DSP state
and logs in %LocalAppData%\Resonalyze. To make a .zip build fully portable,
create an empty file named portable.flag next to Resonalyze.exe. When a newer
release is detected, the version label in the title bar changes to Update
available and breathes slowly between grey and blue so it is noticed on a bar
nobody looks at: installed builds can start an Automatic Update, portable
builds offer a manual download. The pulse pauses while the window is in the
background, stops once you follow the link, and never starts at all when Windows
is set to show no animations (Settings → Accessibility → Visual effects).
Windows SmartScreen note: the builds are not code-signed (certificates are expensive for a free open-source project), so the first launch may show a "Windows protected your PC" dialog. Click More info → Run anyway, or verify the download against the published SHA-256 checksum.
- Band-defined exponential sweep — the low and high frequency it must cover (20 Hz – 20 kHz) plus a per-octave pace, with the transfer estimate gated to the excited band, and impulse-response JSON save/load
- Mandatory loopback for sweep/IR analysis — every IR-based view is derived from the transfer function (harmonics and THD+N stay on the sweep deconvolution); Live Spectrum can additionally run as a reference-free RTA
- Multi-sweep averaging (1–64 runs) as a cross-spectrum estimate with a per-frequency coherence (γ²) curve, and an optional confirm-between-runs pause for spatial averaging
- Analysis views — frequency response, phase, group delay, waterfall, Burst Decay, autocorrelation, harmonic distortion, THD and THD+N, with reliability-anchored phase unwrapping, Fixed/FDW phase windowing, and minimum/excess-phase decomposition
- Calibration — the microphone's 0° profile from
.txt/.cal/.frd/.csvfiles, any number of further named profiles, curves estimated for an off-axis angle from the microphone's geometry (with the uncertainty of that estimate shown), and absolute dB SPL from an acoustic 1 kHz calibrator - Time Alignment — sub-sample delay from the transfer IR, refined by a GCC-PHAT cross-correlation
- Crossover summation prediction — in Frequency Response, the true complex
(vector) sum of two measurements (
Main ⊕ Compare) with Compare delay/polarity controls, plus a sum-loss curve - Virtual DSP — up to eight L/R driver pairs (plus mono channels) through virtual chains: gain, delay, polarity, Butterworth / Linkwitz-Riley / Bessel / Chebyshev crossovers, all-pass and PEQ, with the complex sum, sum loss, phase tracking, junction read-outs, Δ L−R timing, Auto crossover, a stereo-aware Auto delay, a headphone audition, sessions and tuning-sheet export
- Live Spectrum — a real-time loopback transfer function with coherence, or a reference-free RTA in relative dB or dB SPL, with selectable excitation (leakage-free periodic pink, pink, brown/red, white, or Silent for the ambient room) and compensation of the noise's own spectral slope
- Compare a second measurement (file or History) across Time Alignment, Phase, Group Delay, Frequency Response and Impulse Response, and overlays — captured, calculated and target curves with styling, curve math, import/export, saved per-mode state, and a live editing preview
- EQ Wizard — up to 32 PEQ bands toward its own target, from an IR, an overlay slot, a text curve or a Virtual DSP channel handed over for editing (and returned with one click), with Auto Tune, cross-tool import/export and a printable tuning-sheet PDF
- Signal Generator, Measurement History with per-entry working state, a compact Mic/Loop level meter, and four audio backends (MME Compatibility, ASIO, WASAPI Shared and Exclusive) with backend-specific channel routing
Virtual DSP, the EQ Wizard, and Time Alignment are shown in the showcase above; the analysis views:
Frequency Response — smoothing, calibration, an optional dB SPL scale, distortion curves, overlays, targets, and coherence. |
Live Spectrum — a loopback transfer function with coherence and peak hold, or a reference-free RTA in dB SPL, on selectable excitation noise. |
Impulse Response — impulse, envelope (ETC) and step on one timeline, save it as readable JSON, and reuse it across analysis modes without re-measuring. |
Group Delay — timing from the transfer IR, with a millisecond gate, gate offset, and a live impulse-window preview. |
More plots — waterfall, phase, Burst Decay, overlays
To run a release build: Windows 10 or later, working playback and recording devices, a suitable loopback and microphone connection, and optionally an ASIO driver. The self-contained release archives include the .NET runtime.
To build from source: Windows 10 or later, the
.NET 10 SDK — global.json
pins the exact version (rollForward: latestPatch), so an older feature band
fails restore even though it is also .NET 10 — and Visual Studio 2026 with the
.NET desktop development workload, or the .NET CLI.
Use conservative playback levels when connecting physical equipment: start with the output turned down and verify the signal path before measuring.
If you are starting from zero, this is the minimal hardware path — roughly €100–200 total:
- A USB audio interface with at least two inputs (any entry-level two-channel interface with phantom power works). Two inputs matter because every measurement records a loopback reference alongside the microphone — that is what makes the timing analysis absolute. Community-verified so far: Focusrite Scarlett Solo (the developer's own rig).
- An analog measurement microphone (an inexpensive electret measurement mic with an individual calibration file is ideal). A USB measurement mic such as the UMIK-1 will not work — see the FAQ for why.
- Two cables: one to feed the system under test from the interface's output 1, and one short cable from output 2 straight back into input 2 — that is the loopback.
Then, in about ten minutes:
- Wire it up: mic → input 1, output 2 → input 2 (loopback), output 1 → the system's input (in a car: the DSP's aux/optical input, with only the driver under test unmuted).
- Start Resonalyze, open the measurement settings, select the interface, and
assign the input and loopback channels. The measurement will not start
without a loopback — that is by design. Set Measurements to at least
4: the averaged sweeps lift the response out of the cabin's noise floor and produce the coherence curve that tells you which bands to trust. - Turn the playback level well down, place the mic at the listening position, and run the sweeps, watching the input level meter.
- Explore the views: Frequency Response, Time Alignment, Phase, Impulse. Save the impulse response — saved measurements are the raw material for everything else.
- Measure each driver the same way, then open Virtual DSP and let Auto crossover and Auto delay design the tune against the phase-aware predicted sum before you touch the hardware.
The full path with averaging, coherence, spatial averaging, and comparison is described in Measurement Workflow.
Can I use a UMIK-1 or another USB microphone?
No — and it is physics, not stubbornness. Every measurement records a loopback reference next to the microphone signal, and the two streams must share one hardware clock to stay sample-accurate. A USB microphone is its own audio device with its own free-running clock; pairing it with a separate playback/loopback device gives two streams with an unknown run-to-run start offset plus continuous drift, which silently corrupts every timing-sensitive result. This is also why the settings do not offer a separate loopback device.
Why is the loopback mandatory? REW works without one.
The loopback records what actually left the playback chain and exactly when, so every analysis is derived from the mic-vs-loopback transfer function — timing becomes absolute rather than relative to an arbitrary trigger. That absolute time base is what allows separate measurements, taken minutes apart, to be combined later: it is the foundation of the measure-once-tune-at-your-desk workflow, of the complex (vector) sum prediction, and of automatic delay alignment.
Is one microphone position enough to tune a whole car?
At that one point, yes, and exactly: sound pressure sums linearly, so the predicted combination of individually measured drivers is the physics of what the microphone would record — not an approximation. The honest boundaries are the ones any single-point method has: the prediction holds at the microphone position (put it where your head is), in the linear non-clipping regime, with the same playback chain and mic position for every measurement, and at a roughly stable cabin temperature. For frequency-response work you can go further with spatial averaging. And the final judge of a tune is still your ears.
git clone https://github.com/DIMOSUS/Resonalyze.gitThen open source/Resonalyze.sln, or build and run from the command line:
dotnet restore source/Resonalyze.sln
dotnet build source/Resonalyze.sln --configuration Release
dotnet run --project source/Resonalyze.csprojRun all application and deterministic DSP tests with:
dotnet test source/Resonalyze.sln -c Release --filter "Category!=Hardware"That covers Resonalyze.Dsp.Tests (deterministic and synthetic),
Resonalyze.App.Tests (file formats and non-UI application logic against a fake
audio factory) and Resonalyze.Audio.Tests (PCM decoding, capture sessions,
WASAPI configuration). The filter drops the hardware smoke tests, which need real
WASAPI endpoints named through the RESONALYZE_WASAPI_CAPTURE_ENDPOINT_ID and
RESONALYZE_WASAPI_RENDER_ENDPOINT_ID environment variables.
For local performance profiling, build the dedicated Tracy configuration
(dotnet run --project source/Resonalyze.csproj -c Tracy), which defines
TRACY_ENABLE and references Tracy-CSharp. Add instrumentation through
AppProfiler.Zone(...), AppProfiler.FrameMark(...) and
AppProfiler.SetThreadName(...); zones are thread-bound and strictly LIFO, so
never let one span an await.
The Release executable is produced at
source/bin/Release/net10.0-windows/Resonalyze.exe; tagged releases also produce
portable .zip packages for win-x64 and win-arm64, an x64 Setup.exe
installer, and NetSparkle appcast files. The build.yml workflow runs on every
push to main and every pull request: it builds the solution, runs all three
test projects, then proves the release path still works by producing the
single-file publish and compiling installer/Resonalyze.iss. Warnings are
errors.
This workflow covers impulse-response (IR) based analysis: a swept-sine measurement is captured once and then inspected across the frequency-response, phase, group-delay, impulse, waterfall, and burst-decay views. For continuous, real-time analysis without capturing an IR, use Live Spectrum.
- Connect the output of the device under test to the selected input, either directly or through a microphone and a suitable interface.
- Start Resonalyze, open the measurement settings, and select the audio backend,
sample rate, devices or backend-specific input and loopback channels, the
sweep band and pace, playback channel, and
analysis parameters. A loopback reference channel is required. To average
several sweeps, set Measurements above
1; enable Confirm each run to pause before each sweep so you can reposition the microphone. - Start a recording to capture the sweep. With averaging the runs are combined into one transfer IR and a coherence (γ²) curve, debiased by the number of runs: the raw estimate over K averages reads 1/K even for pure noise, so the stored figure maps that null expectation to 0 and stays comparable across run counts.
- Watch the input level meter to confirm microphone level, loopback presence, and headroom before trusting the measurement.
- Select the analysis view you need and adjust smoothing, windows, offsets and display options.
- Capture and compare with overlays: store the current curve in a slot, import a reference from text, or combine slots with curve math. Add a target curve overlay and switch its deviation readout to EQ correction to see how much to dial into an equalizer.
- Pin a second measurement with Compare, use Save to keep the captured impulse response, and History to review recent measurements or reload an older snapshot.
For acoustic measurements, microphone placement and room conditions strongly affect the result. For electrical loopback measurements, make sure the signal levels and impedances are safe for both devices.
The analysis plot, the Time Alignment previews, the EQ Wizard and the Virtual DSP graphs all take the same mouse and keyboard controls, laid out to match REW's graph panel so there is nothing to relearn when you move between the two. (The small previews inside settings panels and dialogs — the impulse window, the gate preview, the history list — are fixed-scale by design and take none of this.)
| Gesture | What it does |
|---|---|
| Wheel | Zooms both axes around the pointer |
| Alt + wheel | The same, in fine steps |
| Shift + wheel | Horizontal axis only |
| Ctrl + wheel | Vertical axis only |
| Wheel over an axis | Zooms that axis alone |
| Wheel over the end of an axis | Moves that one limit, leaving the opposite end where it is |
x / Shift+X |
Zooms the horizontal axis out / in by about two, around the pointer |
y / Shift+Y |
The same for the vertical axis |
| Middle-button drag | Variable zoom: right and left work the horizontal axis, up and down the vertical one |
| Ctrl + right-button drag | Draws a zoom rectangle |
| Right-button drag | Pans |
| The + / − buttons on the graph | Zoom the axis they sit against by about two, click after click; they appear while the pointer is over the plot, and hovering one names the axis it moves |
| Double click | Opens the graph limits dialog |
| Ctrl+Z | Steps back through the zoom-to-area, variable-zoom, zoom-button and fit-to-data moves (a wheel notch is its own undo — scroll it back) |
| Ctrl+Alt+F / Ctrl+Alt+Y | Fit to data / fit the vertical axis to data |
Home or A |
Back to the view's own default scale (also the Defaults button in the limits dialog) |
The graph limits dialog types the same ranges exactly — top, bottom, left and right as numbers, plus Fit to data, Fit Y to data and Defaults, which hands the axes back to the scale the mode chose for them. Use it when two measurements have to be framed identically, for a screenshot or a before/after.
A zoom survives a redraw: changing a setting, running a new measurement or toggling an overlay keeps the range you are looking at. The analysis plot remembers one range per mode, so Frequency Response and Impulse Response do not fight over a scale; the Time Alignment previews keep theirs across a reconfiguration, and the EQ Wizard and Virtual DSP graphs hold theirs until something changes what the axis means (loading a new wizard source, switching the Virtual DSP view between magnitude, phase and impulse).
Axes you have not touched still scale themselves — the dB axis lifts its ceiling for a padded loopback, group delay fits its data — so the automatic framing steps aside only where you took over, and Home hands an axis back to it. Frequency axes pan within 20 Hz – 20 kHz, the band the curves are computed over.
Two Virtual DSP axes answer to their own rule on purpose. Its Phase view spans ±180°, the entire range a wrapped phase can occupy, so the height is locked: there is nothing above or below to travel to, and the limits dialog leaves that axis out rather than offering a range that would only push curves off screen. Its Impulse view zooms and pans in time — the millisecond around each arrival is the part worth reading, and the gate window it opens on is far wider — within that window, which stays the hard limit because the traces hold nothing outside it. Editing the gate re-frames the axis, since that is a new timeline; an ordinary redraw leaves your zoom alone.
The EQ Wizard's right-hand EQ (dB) axis follows the impulse view's rule: it zooms and pans within its nominal range — the boost/cut budget plus the drawn curve, which is its hard limit because the curve holds nothing beyond it — so a correction's fine structure can be read without the whole budget's height. An ordinary redraw keeps your zoom; changing the Min/Max Gain budget (or a curve outgrowing the range) re-frames the axis, since the old view no longer fits what it must show.
The Mode Settings... button opens the current mode's settings in a docked, title-bar-less panel aligned to the plot area, which stays open while the main window has focus and switches automatically when you change modes. Settings apply on the fly, redrawing the analysis while preserving the visible plot range. Each curve-based view groups its plotted curves under a Curves: heading with one checkbox per curve — Primary / HD2–HD4 / THD+N in Frequency Response, or measured / minimum / excess in Phase. Numeric and dropdown settings carry a small R button that resets them to the built-in default, the plot keeps the range you zoomed to (see Graph Zoom and Limits), and the Frequency Response, Phase, Group Delay, Waterfall and Burst panels include a compact impulse-window preview.
The Tools modes (EQ Wizard, Signal Generator, Virtual DSP) do not measure and do not draw the shell's curves: they bring their own sources and controls, so the measurement block on the right — input meters, Start, Record Settings, Save / Load / Compare, History and Mode Settings... — is hidden while one of them is open, and any docked Record Settings or History panel closes with it.
Phase and group-delay analysis run on the loopback transfer impulse response, so they are only drawn when the active record contains one. Both use a millisecond-based fixed gate built from a left Tukey fade, a flat plateau, and a right Tukey fade. A Gate offset positions the end of the left fade inside the analysis frame, and the Auto checkbox (on by default) keeps that offset snapped to the detected start of the impulse response — the band-limited first-arrival front, not the peak — falling back to the transfer-IR peak when the detector cannot get a trustworthy reading. A read-only readout shows the gate's lowest reliable frequency (≈ 1 / gate length).
Phase additionally offers Window: Fixed / FDW. Fixed is the single Tukey gate
across the whole spectrum; FDW builds a bank of time-aligned spectra whose
effective right-side duration follows cycles / frequency, so low frequencies
retain the long window while mid and high frequencies progressively reject the
late reflection tail. FDW cycles selects 4 (strongest suppression), 6 (the
recommended balance), or 8 (more late detail).
The Phase view shows four independently toggled curves: measured phase, minimum phase (the part tied to the magnitude and correctable with EQ), excess phase (measured minus minimum — the all-pass part an equalizer cannot fix), and coherence (γ²) from averaged runs. Detrend removes one constant delay before unwrapping: Auto estimates the slope-based excess delay from the displayed spectrum and shows it in τ (ms), Manual uses the editable value, Off keeps the absolute slope. With Main and Compare together, Auto is resolved once from Main and applied to both, so their real relative delay stays visible as a linear phase difference.
Unwrapped phase uses a reliability-anchored algorithm instead of naive bin-to-bin accumulation: each bin takes the 360° branch closest to a phase predicted from the last trustworthy bin and the running slope. Bins well below the local magnitude envelope — or with low coherence — are still displayed but never trusted as anchors, so deep nulls and masked bands are bridged cleanly, while a stretch too long to bridge honestly is blanked instead of guessed.
Group Delay reads absolute delay referenced to the start of the transfer IR, so a peak well into the impulse response reports its true arrival time, and the curve is computed energy-weighted so near-null bins follow the dominant energy instead of the singularity. FDW is deliberately not applied here: an FDW phase curve is direct-sound-oriented and is not the exact integral of the displayed fixed-gate Group Delay, so selecting Fixed phase restores the compatible pair.
Resonalyze can run measurements through four backends, chosen in the measurement settings dialog:
| Backend | Use it for |
|---|---|
| MME Compatibility | Ordinary Windows playback and recording devices. The most compatible option and the fallback when nothing else works. |
| ASIO | Audio interfaces with a native ASIO driver: lowest latency and arbitrary multi-channel routing. |
| WASAPI Shared | Windows endpoints without an ASIO driver, while other applications keep using the device. The endpoint's own mix format applies, so Windows may resample. |
| WASAPI Exclusive | The same endpoints taken exclusively: no Windows mixer in the path, and the requested sample rate and bit depth reach the hardware unresampled. |
Both WASAPI modes address devices by endpoint id rather than by index, so a chosen device survives reboots and device reordering, and they use the same microphone and loopback channel selection as MME.
The microphone input is the primary measurement channel, and a loopback reference channel is required for every measurement. Both are recorded simultaneously and the main impulse response is derived as a transfer function from the loopback reference to the microphone response, which removes the playback path (DAC, amplifier, output routing) from the analysis. All IR-based views come from this transfer IR; harmonic distortion, THD and THD+N use the ordinary sweep-deconvolution response instead, drawing HD2–HD4 at the excitation frequency (a second-harmonic hump from a 1 kHz drive appears at 1 kHz, not at 2 kHz) so each ends at Nyquist/n. Because those curves sit on the sweep-deconvolution scale while the primary curve is loopback-normalized, their vertical distance is not yet a calibrated distortion percentage.
The exponential sweep is described by the band it must cover: a Low frequency (Hz) and a High frequency (Hz) anywhere between 20 Hz and 20 kHz, plus a Per octave (ms) pace that sets the duration. Measuring a tweeter through a 2 kHz crossover no longer means sweeping from 20 Hz and pinning the top to Nyquist — sweep the band the driver actually plays. Phase alignment is preserved by rounding the band outward to whole start and end cycles, so the achieved range always encloses the one you asked for and the fades live in guard bands outside it; the Actual range line reports what the settings really deliver instead of quietly shortening the sweep at run time, and the transfer estimate is gated to the excited band, because outside it the transfer function is only microphone noise divided by the reference's leakage skirt.
The excitation plays at a fixed −6 dBFS, and there is no control for it. That
is the level the Signal Generator at its default Level, %
of 50 and the Live Spectrum noise already play at, so an
output level dialled in with either still holds when the sweep runs; a full-scale
sine sweep is also the worst case for the converter, which can clip on
reconstruction even when no sample exceeds full scale. The headroom costs 6 dB of
signal-to-noise ratio and nothing else, since the inverse filter carries the
reciprocal scale and the transfer function is scale-invariant.
Save sweep as WAV... writes the sweep the panel currently describes to a 24-bit WAV file — the same band, pace, sample rate, playback channel and level a measurement would play, with a second of silence before and after — for measuring from a source that is not this computer, such as a phone, a head unit or a USB stick in the car. Measurement options otherwise apply as you edit them and touch the audio session only when its identity changed; the audio backend, the device format and its device panel commit together with Apply settings.
Most installs keep a protective high-pass in the external DSP so that nothing — a sweep included — reaches a tweeter or a small midrange below its safe band. That filter is part of what the microphone hears, while the loopback reference is captured before the DSP, so the measured transfer response carries the protection as if it were the driver's own roll-off.
The HPF row tells Resonalyze which filter is in the way: Butterworth (6 dB
steps up to 48 dB/oct) or Linkwitz-Riley (12 / 24 / 48 dB/oct), plus its corner
frequency. The measurement then divides that known magnitude and phase out of
the loopback-referenced transfer impulse response — the equivalent of filtering
the clean reference through the same high-pass before dividing, while the
full-band loopback stays available to the H1 estimator and its coherence.
Inversion has a limit. The compensation is capped at 40 dB of boost, because deeper into the stop band the protection has buried the driver under the noise floor and no arithmetic brings it back. Confidence is full until 6 dB before that ceiling and fades to zero along a raised cosine at it; the same fade multiplies the coherence curve, so a frequency the compensation could not recover reads as untrustworthy instead of as a confident number. Analyses that choose a band to work in read that masked coherence and stay out of the unrecoverable region — Time Alignment's dominant band above all, which would otherwise happily time a driver on boosted noise.
Leave it Off when there is no protection in the chain, or when the loopback is
taken after the DSP — the reference already contains the filter then, and the
division has removed it before this setting could.
Use MME Compatibility for ordinary Windows playback and recording devices.
(Older releases and the settings file call this backend Wave.) You choose the
playback device, the recording device, the sample rate, the playback channel, and
the microphone and loopback input channels (Left / Right, loopback required).
The loopback is captured from a second channel of the same recording device as the microphone, so both signals share one hardware clock and stay sample-accurate; a second input device would put the two streams on independent clocks with an unknown offset plus drift, so Resonalyze does not offer a separate loopback device at all.
Use ASIO for audio interfaces with a native ASIO driver. You choose the
driver, the sample rate, the microphone and loopback input channels (loopback
required), the output channel pair, and the routing within it: Mono sends the
signal to both channels of the pair, Left and Right to one of them, Stereo
to both. Before applying, Resonalyze checks whether the driver supports the
current sample rate, showing its playback latency and a
"supported / not supported" line; a driver already in use by another application
is reported before the measurement starts.
ASIO Control Panel opens the driver's own panel for buffer size or clock source; Test ASIO Inputs captures a short diagnostic snapshot that verifies the microphone and loopback channels are truly separate and not mono-summed by the driver or the interface's control software.
The right-side control column includes a compact two-channel input meter for
Mic and Loop: the bar shows a filtered RMS level, the bright vertical marker
Peak Hold, and the text Peak / RMS in dBFS. After a measurement completes it
retains the final levels from the last valid capture, which makes it easy to spot
missing loopback, a weak microphone level, or overload.
What a level meter cannot show is analog distortion: an input stage driven past its limit distorts long before its digital level reaches full scale, so a loopback reading a comfortable −15 dBFS can still deliver a badly misshapen copy of the sweep. That matters most for the reference, because every analysis is the microphone divided by it — a nonlinear reference produces a wrong answer, not a noisy one, and coherence stays high while it happens. Resonalyze therefore reads each channel's own harmonic content and names the offender when it refuses a measurement:
The LOOPBACK REFERENCE is distorting: its harmonic packets read −8.1 dB relative to the direct one, where the microphone reads −40.6 dB, and it peaked at only −18.1 dBFS, so the input meter had nothing to show.
The fix is to attenuate what reaches the loopback input — a line input instead of an instrument one, a pad in the cable, or a lower playback level — but only as far as it takes to leave the input's linear region, since a pad attenuates the signal and not the input's own noise and a reference driven toward the noise floor pays for it in coherence.
The Live Spectrum mode runs in one of two explicitly chosen Modes.
Transfer is a live, dual-FFT transfer-function analyzer. It plays a continuous excitation signal, uses the configured loopback channel as a reference, and shows the real-time relationship from loopback to microphone, which suppresses input-side content not correlated with the playback signal. Alongside it a coherence curve (γ²) is drawn on a secondary 0-to-1 axis: values near 1 mark trustworthy frequencies, low values flag bands dominated by noise, reflections, or non-linear behavior. The estimate averages in the power domain, and on-screen smoothing is referenced to wall-clock time, so the response stays consistent regardless of overlap and sequence length. The mode needs a loopback reference: without one the choice turns amber and the analyzer runs reference-free anyway, because there is nothing to divide by.
RTA is that reference-free analyzer as a deliberate choice — the microphone's own magnitude spectrum, no transfer function and no coherence. It is what the car workflow actually wants: a moving-microphone average, where there is no reference to divide by. It captures the microphone alone even when a loopback is configured, so the sound card is asked for exactly the channel that is used.
Two checkboxes belong to RTA and are muted in Transfer:
- dB SPL puts the RTA on a true absolute axis (mic level plus the SPL anchor offset), integrated as power per fractional-octave band. A transfer function is a dimensionless ratio with no scalar level under noise excitation, which is why the scale exists here only.
- Slope compensation removes the tilt the excitation itself prints on the
curve, so a flat system reads flat whatever the noise colour: pink otherwise
falls 3 dB per octave on the per-bin dB axis, and on the banded dB SPL display
even flat white noise climbs 3 dB per octave, since a band's power grows with
its width. What is subtracted is the shape the chosen noise really has —
modelled from the generator's own filters, not from a nominal slope, so brown's
leaky integrator and pink's filter bank are compensated to their true response
and not to a straight line their bass does not follow. The curve is pinned at
1 kHz, the plot title says it is compensated, and an overlay captured from it
keeps the compensation.
Silenthas no known excitation spectrum, so the checkbox is unavailable there.
Signal Type selects the excitation: Pink noise (periodic) (the default —
one FFT-length period of exactly pink noise, looped; being periodic with the
analysis block it is measured leakage-free with a rectangular window and
converges almost instantly, so Window is forced to Rectangular and
Overlap to Off), Pink noise (continuous random, −3 dB/octave), Brown
/ red noise (−6 dB/octave, for subwoofer and room-mode work), White noise
(flat energy per hertz), or — in RTA only — Silent, which plays nothing and
measures whatever the microphone hears: the ambient room, or an external source
playing its own material.
Further settings: Sequence Length (the FFT block size), Overlap (Off /
50% / 75%, reclaiming the samples a tapering window attenuates at the block
edges), Smoothing, Window (Hann, Flat Top for amplitude accuracy on
tones, Blackman-Harris for leakage suppression, or Rectangular), and
Averaging (Fast / Medium / Slow time constants, or Infinite, with
Reset Average). The drawn curves are the Main curve, Peak Hold,
Coherence, and RTA (input) — the plain magnitude spectrum of the
microphone alone, drawn beside the transfer function while you are in Transfer
mode, and the whole plot in RTA mode, where it can be captured into an overlay
slot and equalized in the EQ Wizard.
Coherence Limit draws any frequency below the chosen percentage (default
25%) dimmed and dashed, and a processing overload warning appears if the CPU
cannot keep up. Changing anything the capture itself depends on — the mode, the
signal, the window, the FFT length, the overlap — clears the accumulated average
rather than redrawing the previous run's data under the new settings.
Every magnitude-smoothing selector (Frequency Response, Live Spectrum, Fourier Waterfall, Virtual DSP, EQ Wizard, magnitude overlays) also offers Psychoacoustic: smoothing whose width follows frequency — 1/3 octave at and below 100 Hz, narrowing smoothly to 1/6 octave from 1 kHz upward. It shapes magnitude curves only; phase, group-delay and coherence traces fall back to the plain 1/6-octave width, and the Auto delay engine never reads display smoothing.
The History button opens a docked panel with a list of recent snapshots, a
compact frequency-response preview for the selected row, and row tooltips
carrying capture metadata (time, mode, sample rate, duration, channel, peak
index, stored meter levels). Entries come in two kinds: RAM for in-memory
snapshots from the current session, and FILE for saved IR files remembered
across launches; the newest appear at the top, in a stable chronological order.
Double-click a row to load it. Use Save to turn an in-memory snapshot into a regular IR JSON file, Delete to remove an item from history without deleting the file from disk, and New session (reset to defaults) to start clean — all per-mode settings return to their defaults and the current measurement and overlays are cleared, while audio device and routing settings, the history list and saved files are left intact.
Each entry also remembers the working state it was last used with: the active mode, every per-mode setting, and which overlay slots were shown — so switching to another entry and back restores the whole working context, not just the impulse response. Only a small rolling set of unsaved in-memory snapshots is retained.
The Compare button overlays a second measurement on top of the current one, so two responses can be read side by side with the same analysis settings. Choose the reference from a file (Choose file…) or from a History entry; the button then shows its name, and Clear removes it. Compare is applied everywhere it is meaningful, always recomputed with the current mode's settings:
- Time Alignment — the reference envelope is overlaid on the peak preview
with its own markers, and the delay table gains a second block whose every
value shows the delta against the source (for example
1.006 (+0.010)). In Auto band mode the analysis band is then the one the two records SHARE — the overlap of their own dominant bands, labelledshared with Compare— because two arrivals are only comparable where both drivers play, and a band taken from the source alone would make the delta depend on which of the pair was loaded first. Records that barely overlap (a subwoofer against a tweeter) keep the source's own band. - Phase and Group Delay — the reference curves use the identical gate/window and smoothing, drawn dashed and dimmed; Phase Auto detrend is resolved once from Main and reused, preserving their relative delay.
- Frequency Response and Impulse Response — the reference magnitude and impulse are drawn alongside the source (harmonics stay source-only), and with a transfer IR an absolute sample timeline lets the two arrivals be compared.
A reference is only drawn when its sample rate matches the current measurement. The source and Compare curves are also selectable as live operands in a calculated overlay, so their difference can be watched live while you tune the analysis window.
The Impulse Response mode draws the loopback transfer IR from the record start through the peak and on into the tail, so the arrival, the reflections and the decay are one picture — and so two records can be read against one clock.
The traces are built over the whole record, so navigating it is a gesture and not a trip to the settings panel: zoom out to the end of the tail, in to a single sample, with the usual graph controls. Length only frames the view the mode OPENS on — that much tail past the peak — because a deconvolved record is mostly silence and opening on all of it would draw the response as one vertical line.
Three traces share that timeline, each switched on under Curves:
- the impulse response itself;
- the envelope (ETC) — the analytic-signal envelope, where reflections read as separate arrivals instead of as interference in the waveform. ETC smoothing averages it over a chosen duration, centred so nothing moves in time;
- the step response — the running integral of the impulse: what the system would do if the input jumped to a level and stayed there. It is always drawn normalized on an axis of its own (against the impulse peak, or against its own peak when Step against IR peak is cleared), because a record with any low-frequency content integrates into a step many times the impulse peak.
Band filter reads all three traces through a zero-phase band — a full octave or a third of one, centred on any ISO preferred frequency. This is how you see when a band arrives: a full-range impulse buries every band's arrival in one waveform, and the filter is the same raised-cosine bandpass the Time Alignment probe uses, so the view and the delay estimator read the record through the same instrument. Zero phase means the filter delays nothing — at the price of a symmetric ring around each arrival, which is visible in the trace and is why an unfiltered arrival marker stays on the plot beside it. With a band selected the plot title names it and the peak marker becomes the band peak, captioned with how long after the record's arrival that band peaks — the figure the filter exists to produce, since a driver's low band does not arrive when its broadband front does. The caption appears only where the record actually carries the driver's energy at that centre: on the archived cabins a tweeter measured at 63 Hz still had a "band peak", and it landed seconds after the arrival because what peaked there was leakage. It is stated as time and not as a distance on purpose — this delay is the driver's own build-up, not a path through air.
Amplitude scale selects raw Linear sample values — absolute, and therefore comparable between records — or a normalization against the peak, in % of peak or in dB re peak. With a Compare reference loaded, both curves are normalized against the same peak, the main record's: how far one sits below the other is the point of the comparison, and normalizing each to itself would erase exactly that.
Time axis switches the unit between milliseconds and samples; the tracker reads both either way. Time zero chooses where the axis puts its origin — the record start, the estimated first arrival, or the peak. This is a view setting: the measurement is never rewritten, because Time Alignment, the Virtual DSP gate pins and every saved offset are statements about the record's own absolute timeline. When zero sits on an arrival, the tracker also reads the path length that time corresponds to in air. Invert polarity flips the displayed impulse and step the same view-only way.
An overlay captured here stores the record's own coordinates — absolute sample indices and raw levels — and is redrawn under whatever framing the view has later, so it follows the time unit, the time zero, the amplitude scale and the polarity flip instead of staying frozen in the ones it was taken under. Levels are re-normalized against the LIVE record's peak, so how far the snapshot sits below what is being measured now stays readable. Two things cannot be undone that way and travel baked in: the band filter and the ETC smoothing are part of the values. A very long record is stored thinned to its extremes, so zooming an overlay to sample level shows the thinned outline where the live trace shows samples.
Two markers name the instants the rest of the app acts on: the estimated arrival — the same shared figure the Auto gate offsets are anchored on — and the strongest peak, labelled with the record's signal-to-noise ratio whenever the envelope is on screen to measure it against — the same figure Time Alignment reads off that record, because it is computed the same way from the same envelope.
The Time Alignment mode analyzes acoustic delay from the currently active measurement record, for practical loudspeaker, microphone, and channel alignment work where the result has to be more precise than a single audio sample.
It reads the transfer impulse response already stored in the current record, so it works immediately after a sweep captured with loopback or after loading an IR JSON file with transfer-response data. The delay estimator uses a robust two-stage chain: the transfer IR, through an optional raised-cosine bandpass window, gives an analytic-signal envelope whose first arrival and strongest peak are the coarse, polarity-blind anchors; a GCC-PHAT (phase-transform) correlation from the same spectrum then refines each anchor to sub-sample precision wherever its own peak is trustworthy.
The first-arrival search rejects pre-ringing sidelobes by testing every candidate against the analysis kernel's own envelope — an arrival can pre-ring no louder than that allows at a given distance, so a candidate above the ceiling is a genuine arrival no matter how the surroundings look (which keeps weak direct sound alive in reverberant bass), and one at or below it is confirmed as pre-ring by its mirror twin.
It also refuses to read a ripple on the foot of a wave packet as that packet's arrival. A cabin's comb interference leaves small bumps a fraction of a millisecond ahead of a front; they are far too loud to be the transform's own ringing, so the rule above rightly keeps them, yet they are not the front. A candidate must reach a quarter of the strongest envelope level of its own packet — the level the broadband onset calls the onset — or the packet's own front is taken instead. The packet runs one millisecond forward and ends early at a null deep enough (20 dB) to resolve two events, because destructive interference nulls faster than an envelope rises: an earlier arrival separated from what follows by such a null is a separate arrival and keeps its own timing, however strong and however close the next packet is, and the rising edge of a later reflection can never be borrowed to dwarf the direct sound in front of it. A soft direct arrival sitting under a room mode milliseconds later still wins for the same reason, which is what the search depth exists for. Without this two identical drivers in opposite doors could be measured at different points of their fronts — one at its packet peak, the other 20 dB down its own foot — and the level difference lands in the reported delay: on a field pair, 0.31 ms of a 1.45 ms split.
The second stage is what makes the numbers trustworthy. The transfer IR's
spectrum already carries the microphone-to-loopback cross-phase, so whitening it
to unit magnitude over a soft band mask (weighted by coherence where the record
has it) collapses the correlation to a sharp peak at the true broadband delay,
independent of the driver's own magnitude shape. The search runs on peak
magnitude, so a polarity-inverted arrival is located just as reliably. Where the
whitened peak is too weak to trust, the estimate falls back to the envelope's own
interpolated peak instead — and says so: the alignment confidence read-out
gives the normalized height of the whitened correlation peak as Alignment: NN%
and names the method that placed the sub-sample position, GCC-PHAT or
envelope fallback. The payoff is delay estimates such as 87.0 samples or
1.972 ms resolved to a hundredth of a sample.
When the strongest peak lands well after the first arrival — the classic narrowband-subwoofer case — Time Alignment flags it and points you at the first arrival, so a modal or reflected peak is not mistaken for the driver's real timing; the flag requires a real valley (6 dB) between the two peaks, since a low-frequency driver's direct sound can keep rising for milliseconds.
The mode recalculates when you switch into it and as you change the bandpass
settings, and reports signal quality from the analysis envelope and the stored
meter snapshot: a color-coded Excellent / Good / Fair / Poor signal
grade from the recording's SNR; the first-arrival prominence relative to
the strongest peak (a low value means the pick sits on a broad leading edge —
normal for band-limited low-frequency drivers); peak and RMS levels in dBFS; a
CLIP warning; and a FULL SCALE marker for a loopback reference at 0 dBFS.
The measured time, distance, and sample count are clickable: click a result line to copy just the numeric value to the clipboard. With the bandpass window enabled, a frequency-domain preview of the pass band is shown along with the envelope around the detected peak; selecting a Compare reference overlays its envelope there and adds a second delay-table block. Both envelopes are drawn against one reference — the Main record's strongest peak, named in the axis title — because how far a pick sits below its own peak is a figure of the analysis, while the peak is a property of the record: normalizing each curve to its own pick made two records that differ by 4 dB read 19 dB apart.
After a sweep completes, click Save to store the measured impulse-response
data under a timestamped name such as Resonalyze-IR-2026-06-15_14-30-00.json.
Files are indented, human-readable JSON containing the format and schema version,
save time, sample rate and bit depth; the requested and achieved sweep band with
its duration and sample count; the playback channel and measurement mode; the
sweep-deconvolution samples plus the optional loopback transfer-function samples;
optional coherence (γ²) data with the run counts; the stored meter values; an
optional SPL calibration anchor; and embedded
preview frequency-response data for the History panel.
Click Load to open a previously saved response. Resonalyze validates the file
first, rejects files below 44100 Hz, restores the measurement metadata into the
active record, and redraws the current view from the loaded data — without
rewriting the audio-device configuration. Saving and loading are disabled while a
measurement is running. The current file format identifier is
resonalyze-impulse-response, version 7.
Load also accepts a .wav file — a recording of the sweep made outside
Resonalyze, by a phone, a handheld recorder or a DAW, while the excitation was
played from something else (typically the file written by
Save sweep as WAV...). The recording is deconvolved
with the inverse filter of the sweep the current settings describe, and the
transfer function is estimated against that same sweep standing in for the
loopback reference. If the file has more than one channel, the one whose content
actually matches the sweep is measured — not the loudest, because a dead input's
hum can easily out-measure a quiet microphone.
The recording may be far longer than the sweep. The excitation is found by matching the sweep against the recording rather than by looking for something loud, which reaches much further down since matching concentrates the whole excitation into one peak (~46 dB for a two-second sweep); only the excitation plus 0.5 s before and 2 s of decay after is then analyzed. Player and recorder run on their own clocks, so the sweep comes back slightly stretched or squeezed: Resonalyze finds the stretch that sharpens the arrival most and rebuilds the reference at that rate. The settings must match the sweep the recording was made from, and Resonalyze checks that rather than trusting it — against the wrong band or pace a recording deconvolves into a smear instead of an arrival, and arrival sharpness tells the two apart. A wrong sample rate, a sweep running past the end of the recording, a clipped take, and one that does not deconvolve credibly are all refused, leaving the measurement on screen untouched.
Two things such a measurement cannot carry: absolute time, since where the
arrival landed was decided by when the recorder was started, and a
dB SPL anchor, since the recording chain's gain
is unknown. Because the origin means nothing it is chosen rather than inherited —
the impulse response is rigidly shifted so the arrival lands at 10 ms, preserving
every delay inside the measurement. That the timing is local travels with the
measurement into the saved .json and the history, so everything comparing one
arrival against another refuses it by name: Time Alignment
declines it as a source and as a Compare partner, and
Virtual DSP will not sum it with another measurement.
Each supported overlay view provides twelve universal slots, each holding one of
three kinds: a Captured snapshot of a curve currently on the plot; an
Operation between two operands, each a live plot curve or a captured slot
(A - B, B - A, A + B, (A + B) / 2, |A - B|, or a frequency blend), plus
the complex (vector) sum described below; or a parametric Target compared
against a source.
Slots are stored automatically as human-readable JSON under the application data
directory, as overlays/<AnalysisMode>/overlay-01.json. The numbered button
opens a menu to Capture curve, Import from text, Export to text, or
switch the slot to a Calculated overlay or Target; the checkbox shows or
hides it, the numeric control applies a vertical offset, and ⚙ Settings…
opens its dialog. A live-curve operand re-reads the plot on every rebuild, so a
calculation over it — for example the difference between the source and a Compare
curve — updates live as the analysis settings change.
Captured overlay settings cover a name, line color, thickness, style and opacity,
optional 1/48 … 1/3 octave smoothing, and a Clear action for that slot
alone; calculated overlays add the operands, the operation, optional
amplitude-space math for dB views, and independent smoothing applied afterwards.
In Phase Response the difference operations are phase-aware: a wrapped
operand makes the difference take the shortest angular distance so it never jumps
by ±360°. Overlay JSON always stores the unsmoothed source points, so changing
smoothing is lossless, and operations are applied to the displayed Y values after
source offsets — so addition and averaging on a decibel plot are arithmetic on dB
coordinates, not physical summation of acoustic power.
A Target overlay compares a source against a parametric target shape and draws two curves from the one slot: the target itself and the deviation (source minus target), plus an optional shaded tolerance band (±dB). The source is either a captured slot or the current measurement — the Frequency Response curve, or the Live Spectrum trace, which target and deviation follow frame by frame.
The shape is built from four editable terms — an overall tilt around a 1 kHz
pivot, a bass shelf, a treble shelf, and a presence bump/dip — with
editable presets: Flat, Room (gentle), Room (Harman-style), Warm, Car,
Car (mild), Car (bass), House / bass boost, X-curve (cinema), Smiley,
BBC dip, Custom. The three car presets share one in-car shape — a bass shelf
over a flat 400 Hz…5 kHz band, then a gentle rolloff reaching ≈3 dB by 20 kHz —
and differ only in how much bass they lift (+6, +9 or +12 dB); a new target
overlay opens on Car. X-curve (cinema) follows ISO 2969 / SMPTE ST 202 —
flat to 2 kHz, then ≈-3 dB/oct. The deviation curve is Deviation
(measurement − target), EQ correction (target − measurement, the gain to
dial into an equalizer), or None. Target overlays are available in Frequency
Response and Live Spectrum.
Import from text loads a captured overlay from a plain-text file of X Y
pairs (for example, 123.4 -5.5), one per line, parsed leniently: any separator,
extra columns ignored, non-numeric lines skipped. Export to text writes the
slot's current curve in the same format; for a Target slot, Export deviation
writes the deviation or EQ-correction curve. Exported files open with a commented
# resonalyze-curve header recording what the curve is — the analysis it came
from, its role, and the sample rate where one applies. Foreign files without the
header still import as before; the header only lets Resonalyze recognize its own
curves on the way back in, so that the EQ Wizard can tell a
measured response from an EQ-correction curve that must never be equalized as if
it were one.
In Frequency Response, a calculated overlay can compute the complex (vector)
sum of the Main and Compare transfer impulse responses (Main ⊕ Compare).
Both share the same sample-0 time reference, so summing them sample-by-sample and
taking the magnitude gives the physically correct summed response of two
sources — accounting for relative delay, polarity, and phase, unlike arithmetic
on dB magnitudes. Two Compare-side controls make it a DSP-style alignment tool:
Time offset, a fractional-sample delay applied to the Compare IR, and
Invert polarity, both updating the summed curve live.
A companion sum loss operation (complex − magnitude) plots the difference
against a phase-blind magnitude addition. By the triangle inequality it is always
≤ 0 dB: zero where the two sources are perfectly in phase, dropping into deep
negatives toward cancellation. Only the complex side moves as you tune the offset
and polarity, so the curve rises back toward 0 dB as you bring the sources into
phase — a direct read-out of the summation loss you are dialing out.
The captured, calculated, and target settings dialogs preview their result on the
plot while you edit, and Cancel (or Esc) restores the previous state.
Overlay files are separated by analysis mode and restored automatically; all
slots use one file format, resonalyze-overlay, version 5 (older schema
versions are intentionally not loaded). Overlays are available in the Impulse
Response, Frequency Response, Phase Response, Group Delay, paused Live Spectrum,
and Autocorrelation views, with a Show all / Hide all pair.
The EQ Wizard (under the Tools tab) designs a parametric equalizer — up to 32 peaking (PK) bands plus a preamp — that moves a measured response toward a target. It owns its own target curve, edited through the same dialog the Target overlays use but stored with the wizard's own settings, so tuning here never disturbs your overlay slots.
The Source… button picks the curve to tune, and it does not have to be an impulse response: an impulse response from file or history, a curve from an overlay slot (a snapshot, with no live link back), or a curve from a text file. The case this was built for is a moving-microphone RTA in dB SPL: park the Live Spectrum RTA on a car's listening area, capture it into an overlay slot, and equalize that — such a curve has no impulse response and no coherence behind it, and its datum is absolute rather than relative. Only measured responses can enter: a harmonic, THD, phase, deviation, EQ-correction, target or calculated curve is refused, and imported curves carry their own Calibration choice, because a curve captured through a calibrated RTA must not be calibrated a second time.
The plot shows, on shared frequency/dB axes: Source (with optional extra smoothing), Target, Source + EQ, the EQ filter response itself (on its own right-hand dB axis), and a shaded error fill. Click a band card to overlay that band's contribution as a dashed curve. Each card carries its frequency, Q, and gain, and the panel adds a Target Level, a Gain (preamp), a Bands count, source Smoothing, and Bypass. The Target Level is the user's knob alone — loading a source never moves it, so a deliberately placed target survives every source switch (an absolute dB SPL curve simply needs the level dialed to its datum once). The one exception carries rather than guesses: a Virtual DSP handoff brings that panel's own target level along, below.
A Virtual DSP channel's PEQ row opens the wizard on that channel directly — Edit in EQ Wizard on its menu, taking the side the panel is showing (a mono pair hands over its single set). The wizard then shows the very curve the user just left: the measurement through the channel's DSP chain with the PEQ bypassed — the one stage under edit — under the same steady-state window and microphone calibration the Virtual DSP magnitude view uses, the smoothing selector starting on that panel's value. Smoothing is the one thing that then goes its own way: it is a reading width, not part of the tune — the filters do what they do at any smoothing — so turning it here changes what you look at and what Auto Tune fits against, without making the resulting bank belong to a different channel. Edit raw in EQ Wizard hands over the raw measurement instead — the panel's Raw curve — for tuning the driver itself irrespective of the chain.
One case parts from "the curve you just left", deliberately: a bypassed block contributes its raw signal, so the plot is not drawing that chain at all. The handoff still opens on the chain, because that is what the PEQ will live in the moment bypass comes off — a bank tuned against a crossover-less curve would be wrong for the setup. The menu item says so before the trip (it reads chain — block is bypassed), and so does the source description in the wizard.
That identity extends to the corrected curve: Source + EQ is not the bare curve with the filters' ideal magnitude added on top, the way an equalizer normally previews itself. The wizard runs the whole chain — the bank being edited included — through one pass and windows the result, exactly as the panel does for a channel carrying that PEQ, so the preview is the panel's own arithmetic rather than an approximation of it. (A window does not commute with a filter; the steady-state window is long enough that the two would rarely part visibly, but the honest path holds by construction, not by luck.) The Tuning results figures are measured against that same curve. It costs a pair of transforms per edit, so it is computed off the UI thread and the last finished curve stays on screen while the next one runs.
The channel's bands and preamp seed the filter bank as one undo step (the bank is the wizard's single global one, so Ctrl+Z is the way back to what it held), and the From / To window lands on the channel's crossover corners — beyond them the chain is rolling the driver off on purpose, and a fit would chase the slope (a raw edit, or a channel with no crossover, leaves the window alone). The Target Level arrives from the Virtual DSP panel verbatim: the handoff curve is rendered in that plot's own dB frame, so one target means one height too — the curve hangs exactly where it hung a click ago. The Calibration selector comes up pinned to the Virtual DSP panel's choice and disabled: a PEQ fitted under one correction and summed under another would break the identity above, so the correction is changed where it lives. The wizard's standing calibration preference for impulse responses survives untouched.
Return PEQ to Virtual DSP — visible only during such a session — sends the finished bank (bands and preamp) back to the channel side it came from, named "EQ Wizard" in its read-out, and switches back to the Virtual DSP tab. The address is remembered from the handoff, so flipping the L/R selector while editing does not misdeliver the result. Back without applying beside it leaves the same way with nothing written: the channel keeps the PEQ it had, and the wizard keeps the edits — exportable, or one Ctrl+Z chain back to the pre-handoff bank. (A plain tab switch, by contrast, keeps the session open for coming back.) Loading any other source ends the session and hides both buttons.
The Target Level travels back with the bank. It is your knob in the wizard, the bank's preamp is fitted against wherever you put it, and returning the filters without it would realize a tune aimed at a height the panel does not have.
A return is refused — with the filters kept, ready for an export or a fresh edit — when what the bank was tuned against has changed since: the channel removed or replaced by another project, that side given a different measurement, its PEQ loaded or cleared from the panel meanwhile, the gate moved, the microphone calibration or the panel's own target level changed, the pair switched between stereo and mono (which moves where the settings live), or any change to the chain the curve was built through. Calibration is on that list for the same reason the wizard locks its own calibration selector during a session — a bank fitted under one correction and summed under another is not the same bank, and the Virtual DSP panel's own selector is a tab switch away.
A polarity flip is the one exception, and the only one: it is −1 at every frequency, so it changes neither the shape the bank corrects nor the level it was fitted against. The rest of the chain does, measured rather than assumed — the crossover bends the curve outright, a gain slides it against the absolute target the bank's preamp was fitted to, and a delay or an all-pass moves what the analysis window catches (at 192 kHz, where the window is at its shortest, by as much as 1.7 and 4.8 dB at the extremes the controls allow).
Auto Tune fits the whole EQ automatically: it works on the error between the target and the (smoothed) source, sets a preamp for the broadband level, then adds peaking bands greedily where the residual error is largest, choosing each band's frequency, gain, and the Q that reduces the error the most. It chooses the band count itself, up to the Max Filters limit (4–32), while a cumulative-boost cap and minimum band spacing keep it from stacking maxed-out bands where the response simply cannot be corrected.
Cuts only (on by default) is the safe choice for a car tune: a boost cannot fill a reflective cabin's interference null — it just burns amplifier headroom on a dip that shifts the moment the microphone moves. Unticking it lets Auto Tune boost where boosting is trustworthy: high measured coherence and not inside a narrow, deep null, still obeying the Max Gain and total-gain limits. A From / To window limits where bands are placed and bounds the error metrics in the colour-coded Tuning results panel, which reports RMS error and Max error between Source + EQ and Target, Filters used, Peak boost and Peak cut, and Headroom (red when the EQ nets a boost that could clip).
PEQ profiles move both ways for Equalizer APO, REW filter settings, Generic CSV,
EasyEffects (JSON), CamillaDSP (YAML) and the Audiotec-Fischer "Full EQ (30
bands)" bank the HELIX / MATCH / BRAX DSP PC-Tool imports per channel (the same
tab-separated block REW exports for that equaliser: PK plus the LS_Q / HS_Q
shelves plus REW's Modal rows, always 30 slots — a bank has no place for the
preamp, so it is not written and the wizard tells you which channel gain to enter
in the PC-Tool instead), and export-only for miniDSP biquads (RBJ coefficients at
44.1 / 48 / 96 kHz) and GraphicEQ (Wavelet / JamesDSP). Import is deliberately
lenient: comments, blank lines, disabled (OFF) filters, non-peaking types, and
malformed entries are skipped rather than rejected. The one exception is a
fixed-layout device bank: the Audiotec-Fischer file is the channel's 30-slot
table, so a truncated or renumbered one is refused outright rather than imported
as an empty bank over the EQ you have — and so is one whose enabled slot claims a
filter that cannot be read, since in a fixed table that band would simply go
missing from the tune (None, the all-pass slots and Enabled False rows remain
ordinary empty slots).
Export as tuning sheet produces a phone-friendly PDF for reading next to the
car: the banner, a title, the date and fit range, an EQ preview graph with the
fit window shaded, the tuning statistics, the preamp, and one card per filter.
Processors do not agree on what the Q of a peaking band means, and the
disagreement is invisible until you cut deep. Every convention states the
bandwidth between the half-gain points as BW = m · Fc / Q and differs only in
the multiplier:
| Convention | Bandwidth at half gain | Behaviour | Seen on |
|---|---|---|---|
| RBJ | Fc / Q |
Independent of gain | Equalizer APO, CamillaDSP, REW Generic/Extended, Audiotec Fischer (HELIX / MATCH / BRAX), Audison/Hertz, Mosconi, miniDSP |
| Symmetric (Zölzer/DAFX) | sqrt(|gain|) · Fc / Q |
Widens as the band deepens, boost and cut alike | AMP Panacea, Behringer DCX2496, Rockford Fosgate 3Sixty.3, Hypex Input EQ, rePhase, Crown USM810 |
| Classic | sqrt(gain) · Fc / Q |
Asymmetric — boost wider, cut narrower | JL Audio TwK-88 |
Resonalyze fits, plots and exports RBJ filters throughout. Hand a Q of 5.8 at −15 dB to a Symmetric processor and it realizes a band over twice as wide. The DSP Q selector states which convention the processor being tuned uses; it moves the Q printed on the EQ Wizard's tuning sheets (which name the convention they were written for) and nothing else — the fit, the curve on screen and the exported profiles stay RBJ. Virtual DSP asks for the convention as it exports, pre-selected from this selector, because a crossover sheet is often written for a different processor than the one the wizard was last pointed at. The conventions are exactly reconcilable:
Q_symmetric = Q_rbj × 10^( |gain| / 40) ±3 dB ×1.19 ±12 dB ×2.00 ±15 dB ×2.37
Q_classic = Q_rbj × 10^( gain / 40) +12 dB ×2.00 −12 dB ×0.50
The lists follow REW's equaliser reference, and are conventions of a model, not of a manufacturer or a chip — JL Audio's TwK-88 and VXi disagree behind the same tuning software. If your processor is not listed, measure it: set one band to Fc 1 kHz and Q 4, at +12 dB and then −12 dB, and read the bandwidth between the ±6 dB points off a sweep. RBJ gives ~250 Hz both times, Symmetric ~499 Hz both times, Classic ~499 Hz and ~125 Hz.
The Signal Generator (under the Tools tab) plays a continuous test signal through the current playback device, independent of any measurement — handy for setting output levels, checking channel routing and polarity, exercising a loudspeaker, or feeding an external analyzer.
Signal type offers the same excitation options as Live Spectrum plus a
Sine tone, Duration, s sets how long it plays, and Level, % scales
its amplitude — the default 50 is −6 dBFS, exactly the level a
measurement sweep plays at, so setting the output
level here transfers to the measurement. The generator reuses the audio
configuration from Record Settings and displays the resolved settings before
you press Play.
The Virtual DSP (under the Tools tab) is the summation-prediction workflow taken to its conclusion: measure each driver once, then design the whole DSP setup virtually. Channels (A, B, C, …) are stereo L/R pairs, each side picking its own measurement and running its own chain. L / R radios switch which side the controls edit, L→R / R→L copy chain settings across sides (a dialog picks the channels and which parts travel — see below), and a Mono checkbox turns a pair into a single shared driver — the typical one-subwoofer car layout — feeding both sides' sums. The setup grows from two up to eight pairs, and +/− folds a block down to its header. Every channel in a project must share one sample rate.
The source button's menu also carries Open in analyzers: it loads that side's own measurement into the analysis modes and lands on Frequency Response, so the driver this channel is tuned on can be inspected with the full toolset — impulse, phase, group delay, waterfall, overlays — and then left again. A history-backed source restores the entry exactly as the History window would (its saved working state included); a file-backed one loads as the Load button does. The entry is greyed out when neither the history entry nor the file behind the channel resolves any more.
Each channel runs through:
- Gain (dB) — relative levels are only honest when the measurements share one playback chain; compensate any difference here
- Delay (ms) with a live mm read-out — the ruler check against the physical driver offset (343 m/s)
- Invert — the DSP polarity switch
- Crossover — Off, low-pass, high-pass, or band-pass; each edge picks Butterworth (6–48 dB/oct), Linkwitz-Riley (12/24/48 dB/oct), Bessel (6–48 dB/oct, near-constant group delay), or Chebyshev (6–48 dB/oct, with a selectable passband ripple) with its own corner
- All-pass — 1st order (180° of phase swing) or 2nd order (360°, with a Q setting how abruptly it turns). Only phase moves, which makes it the tool for lining drivers up where a delay and a polarity flip are both too blunt — a sub-to-midbass hand-off at 60–100 Hz is the classic case — with a live read-out of the group delay it adds (≈ 4Q/ω₀)
- PEQ — one button, five doors: Load from file… (any format the EQ Wizard imports), Save to file… (any format it exports, plus a tuning-sheet PDF), Edit in EQ Wizard and Edit raw in EQ Wizard (the handoff that opens the wizard on this channel's own curve and brings the result back), and Clear. With the handoff there, a whole tune can be built between these two panels without a file in between — so Save is where it leaves for the hardware, going out through the same formats, shelf/preamp rules and warnings the wizard's own export uses. The sheet states the channel's passband when it has a crossover
- Mute and Bypass — Mute removes a channel from the plots, sum, loss metric and Auto delay; Bypass keeps it in the sum but feeds its raw measured signal, for an A/B against the processed result (Auto delay refuses to run while any participant is bypassed). Both belong to the BLOCK: they are shared by its two sides, so muting a driver mutes the pair rather than half of it
- IR polarity — a measured Normal / Inverted / Unknown indicator read from the transfer IR, independent of the virtual polarity switch
L→R / R→L ask before they act: a dialog lists the stereo pairs (mono pairs have a single settings set, so they never appear) and the parts of the chain to carry over — Gain, Delay, Invert, Crossover, All-pass and PEQ. The crossover and the PEQ are ticked by default, because the magnitude shape describes the driver. Everything that aligns a side against its own level and geometry starts unticked — gain, delay, polarity and the all-pass, which belongs with them precisely because it is the tool for a junction a delay and a polarity flip cannot fix, and that junction is the side's own. Sources are never copied: every side keeps its own measurement. Mute, Bypass and the two curve toggles are absent from the list because they are shared by the two sides already — there is nothing to copy.
Because every stage is linear and the measurements are loopback-referenced transfer IRs, multiplying each measurement by its chain and summing the results as complex responses predicts the linear response the microphone would capture after dialing those settings into the hardware. The filters are evaluated as the digital biquad cascades a real DSP runs, so the prediction matches miniDSP-class hardware up to Nyquist, not just an analog textbook curve.
The acoustic plot shows raw and processed curves per channel for the active side (the two per-channel curve checkboxes belong to the block, so a side switch redraws the same curves from the other side's measurement), the complex Sum, the opposite side's Sum as a dashed translucent curve, and the Sum loss curve, with a Phase view toggle and a Sum loss read-out (avg / dip per junction plus a total). Its Gate... dialog exposes Fixed / FDW, 4 / 6 / 8 cycles, and Off / Auto / Manual detrend alongside the IR preview, Tukey controls, and gate offset. Where the gate SITS belongs to the side you are viewing, since the two sides' drivers sit at different distances; how the phase is READ stays project-wide, because two sides read through different windows could not be compared.
The gate's durations shape the phase and impulse views only. The magnitude view — channels, Sum, Sum loss and the read-out built from them — deliberately reads a long fixed steady-state window (~680 ms, clamped to 32768 samples at high rates) that only takes the gate's OFFSET, saying where it opens. The two views answer different questions: phase is timed on the direct sound, where cutting before the first reflection is the point, while tonal balance is what the ear hears with the cabin — and a junction-length gate cannot even contain a bass EQ band's own ringing, so under it a Q 5 cut at 100 Hz would draw at a fraction of its real depth. One window definition serves every magnitude curve here and in the EQ Wizard, which is what keeps the two tools showing the same curve for the same channel. Every magnitude window — here, in the wizard and in Frequency Response mode — opens on the response's detected start, the same band-limited first-arrival front the phase gate's Auto offset snaps to, never on the IR peak: a woofer's peak trails its own onset by several milliseconds of group delay, so a peak-anchored window would open after the response has begun and read the record minus its direct arrival.
A Target checkbox draws the EQ target over the prediction: the SAME target the EQ Wizard equalizes towards, shaped from either place through the same Target... dialog, so the tool that predicts the sum and the tool that corrects it aim at one curve rather than at two that drifted apart. These curves are transfer-function dB with no absolute reference, so the target has no level of its own here — the dB box beside the checkbox says where it hangs. The session stores both: that level, which belongs to this plot's dB reference and so stays put when the shape is retuned, and the target itself — the whole custom shape rather than a preset name, because a preset's numbers can change between versions while a session has to open aiming at the curve it was tuned against. Loading a session therefore sets the app's target to the one it carries, which is the same single target the EQ Wizard shows; a session written before targets were stored carries none, keeps yours, and starts carrying it. A target is a magnitude shape in dB, so it is offered on the Magnitude view only — and every curve toggle follows that rule, muted on the views that cannot draw its curve. The Sum keeps a separate answer per view, since the phase plot is usually one trace denser than you want it while the magnitude plot is where the sum is the whole point.
A gate that opens after a driver has already arrived is refused rather than worked around. The panel judges the placement against every enabled channel and, when one falls outside the window, shows an amber note in the top right corner — above the sum-loss read-out it invalidates — naming the side, the offset, and which channels are cut, with the whole arithmetic (the plateau, the fade-out, each channel's arrival and its leading-edge loss) in the tooltip. A curve gated that way is the reverberant tail rather than the driver, and the sum-loss read-out built from it describes nothing real, so Auto delay and Auto crossover decline to run until the gate is moved: an alignment computed through such a window would optimize the room's answer, not the loudspeaker.
A second plot shows each DSP chain's own magnitude and phase (without the driver) — or, on its Corr mode, one adjacent pair's band-limited GCC-PHAT whitened cross-correlation together with its direct twin, the same comb read on the drivers' direct sound alone, plus the junction's prior-free acoustic score for both polarities. That score is the acoustics alone, while the searches also weigh the arrival prior and the lobe/onset/scene locks, so the gap between the solid marker (the current alignment) and the dashed one (the envelope-arrival estimate) is that trade, drawn. The direct twin is the engine's polarity witness: where the summation score is too close to call between a lag and its inverted rival, the wavefronts within a period or two of the front decide it, because that is the part of the record the drivers made and the room had not yet answered.
A Junction phase block reads each adjacent pair's steady-state cross-phase in
a time-sized window (~0.68 s of the processed IR) — the regime sustained program
material actually sums in, deliberately not the direct-sound phase, because the
room adds several milliseconds of apparent group delay down low. Per junction it
shows φfc, the lower channel's phase minus the upper at the crossover (≈0°
means phase-aligned; ±180° does not by itself call for a flip, since an inverted
channel and a half-period delay are identical at fc); fix ms, the extra delay
on the lower channel that would maximize the overlap-band phase score, with i
recommending a flip and ~ warning that a flip nearly ties; and lobe, how
decisively that delay beats the nearest same-polarity rival. A Δ L−R block
below reports each pair's inter-side state — the two sides' band-limited envelope
arrivals with their difference (positive means the right side leads, the scene
offset's convention), plus a Level Δ L−R row for the by-ear gain trim that
finishes the centering.
Editing a chain recomputes the prediction on a background task, so dragging a value stays responsive with several channels loaded. The Mic cal selector applies one of your configured microphone corrections to the magnitude curves; it defaults to Off because the measurements are loopback-referenced. The session stores WHICH calibration it was tuned with; opening it on a machine that has no such entry says so once and draws the curves uncalibrated, because the other machine's file describes its microphone, not yours.
- Auto crossover... estimates each channel's usable band and driver type (subwoofer, woofer, midbass, midrange, tweeter), asks which filter families to allow, the crossover-frequency window, and whether the two sides of a junction may take independent slopes, then searches frequency, family and slope to flatten the summed magnitude — penalizing wide band overlap and keeping a practical minimum slope, so it lands on a tight, engineer-sensible split rather than shallow filters that only look flat by overlapping widely. The gains follow a car target curve rather than a flat sum: midrange and tweeter are levelled to each other, the lowest bass driver anchors the bass at a chosen elevation over that reference (Bass level over mid/treble, capped at the measured elevation), and the rest are fit onto that slope cut-only, so the result is headroom-safe. Handovers land on human-friendly frequencies, stay in the sensible range for the two driver types, and may only use a slope whose peak group delay stays within 10 ms — fine at a 250 Hz woofer/mid handover (~5 ms), excluded at a 75 Hz sub/woofer one (~17 ms). Heuristics also penalize junctions in the ear's most sensitive band (2–4 kHz), cross two drivers sharing a wide band low (except the subwoofer, nudged UP toward ~80 Hz), and make a low tweeter handover earn its slope against the driver's resonance. Apply then expands ~50 near-optimal variants and re-ranks them by the junction loss actually achievable after the best per-junction delay.
- Auto delay aligns in two stages: band-limited first arrivals, refined by a
GCC-PHAT cross-correlation whose dominant extremum of either polarity seeds the
junction (an inverted junction — a subwoofer against its midbass is the classic
— seeds from the trough, with the polarity decision left to the sum search);
then a fractional-delay search minimizing the sum-loss metric at each junction,
through a direct-sound window so late room reflections do not steer it. That
window is the junction's own: it opens on the earliest front of the two
channels being joined, read in their shared band (never later than the peak, and
falling back to it where the band carries no measurable arrival), and it is
sized by that band rather than by a fixed span — a 60 Hz handover needs
milliseconds a tweeter pair does not. The displayed curves are anchored by the
same rule and differ only in span. The window also travels with its channel
as the search shifts it, because a fixed window over moving content measures the
window instead of the sum. At mid/tweeter-class junctions the search is
additionally locked to the drivers' broadband IR onsets, so the summation
comb can fine-tune only within the physically correct lobe. Candidates are
scored by in-band average loss and the depth of the deepest smoothed notch,
and weighed against an arrival-based prior, so the search does not add delay or
flip polarity without a real improvement. If the resulting delays span more
than ~10 ms — usually one channel's crossover having excessive group delay — a
banner flags the lagging driver.
With stereo pairs, Auto delay tunes both sides in one run, and an
LHD / RHD toggle says which seat you are tuning for. The driver's side is
the reference: it aligns first, the top pair is bridged to it by band-limited
envelope arrivals, and the far side descends junction by junction. The scene
offset is entered as a non-negative magnitude — how far the far side leads —
so switching LHD/RHD never means re-entering a sign, and the level tilt is
entered the same way, as a cut on the near side. The gain balance itself is
off by default: a run writes delays and polarity and leaves every level
alone until you tick Balance channel gains (cut-only). Pairs whose shared band
reaches the localization region are pinned to the scene, because the image
outranks the handover there; a final scene-preserving pass may then shift both
sides of a pair by one shared delta to recover what the pin cost.
A run only proposes: the dialog answers with a report and nothing is written
until Apply. It carries a row per channel — a value the run changes reads
before -> after, one it leaves alonevalue (kept)— over a summary naming the channels each kind of change lands on and the predicted sum loss: the sum loss averaged over the crossover window, before and after, per side in a stereo run, with what the proposal buys (or costs) spelled out. Every delay also carries a confidence — how decisively the measurement supported that pick, withrefthe anchor the others align to andlockeda pick its onset/scene constraint made rather than the acoustics — and aLOWone is named in a warning line, the margin behind it in the notes under the table. - Capture to overlay saves the predicted sum as a Captured overlay in Frequency Response — compare it against real measurements and target curves, or feed it onward to the EQ Wizard.
- Audition track… renders a music file (wav/mp3/flac/m4a and friends) through the tune into a stereo WAV: each program channel is convolved with the summed processed response of its side, with the microphone calibration optionally baked in and one shared normalization gain so the L/R balance survives. Subtract cabin optionally removes a typical body-style cabin transfer function (the pressure-zone bass rise reaching +15…+27 dB at 20 Hz), level-matched so an A/B differs in tone, not loudness: the raw render reproduces the in-car bass rise as headphone boom the in-car listener never perceives, while the subtracted one leaves this car's own deviation audible. Listen through headphones only — it is a stereo auralization of the two sides, not a binaural head simulation.
- Export… writes the whole setup as a tuning sheet (printable PDF or plain text): for every side of every pair (a mono pair prints once) the gain, delay in ms and mm, polarity, crossover filters, the all-pass stage, and PEQ bands. It asks first which Q convention the PEQ columns should be stated in — the processor being tuned here is not necessarily the one the EQ Wizard's DSP Q selector was set for, so that selector only pre-selects the answer (as does the previous export in the same session). The chooser carries a crib that follows the selection: what the convention does to a band's width, and which processors are known to read Q that way. Save session... / Load session... export and import the complete session JSON for sharing or archiving.
The tool's autosaved state persists in tools/virtual-crossover.json and
survives restarts. Accuracy holds within the usual physics: one microphone
position, the same playback chain for every measurement, and the linear
(non-clipping) regime.
Resonalyze applies a microphone (or measurement-chain) frequency-response
correction during logarithmic resampling. In Record Settings, Mic
calibration 0° browses to the microphone's own on-axis correction file. Files
are read leniently in the common plain-text formats (.txt, .cal, .frd,
.csv): frequency level pairs, with comments, headers, a decimal comma,
various delimiters, and extra columns all handled.
Beside it, More calibrations → Manage... holds any number of further calibrations, each with a name of your choosing:
- a file — a second microphone, a different capsule, another chain;
- an angle — a curve estimated for an angle of incidence between 0° and
90°, derived from the 0° file (or from another file entry) plus the geometry
of your microphone: the outer diameter of its front and whether the protection
grid is fitted. The estimate reads the published GRAS free-field corrections as
measured diffraction of known geometries, takes only the change with angle,
scales each reference's frequency axis by the diameter ratio (diffraction
follows
ka = πdf/c), and reports the median of the matching references with their spread as the uncertainty — which for half-inch constructions reaches 2 dB at 20 kHz. The dialog draws that band and states it in words: an angle entry is an estimate from geometry, never a measurement of your microphone off axis. One microphone is modelled from its own measured behaviour instead — the Sonarworks XREF 20, whose 90° difference a generic 12.7 mm estimate misses by up to 2.2 dB.
An angle entry stores the recipe rather than the points, so correcting or replacing the 0° file updates every angle derived from it. Entries are edited on a working copy and applied when the dialog is accepted; angle entries can only be derived from file-backed ones, so an estimate is never built on an estimate.
The views that read a magnitude — Frequency Response, Live Spectrum, the
EQ Wizard and Virtual DSP — each pick one of them (or Off) in their
own selector; Phase and Group Delay read timing rather than level and apply no
correction at all. A selection whose file went missing, or whose entry was
deleted, stays selected and is marked rather than being silently rewritten to
Off. For a source checkout, a legacy source/calibration.txt beside the
executable is still honored as the 0° calibration.
The microphone calibration above corrects the response shape; an SPL calibration anchors its absolute level, so the Frequency Response and the Live Spectrum RTA can be read directly in dB SPL.
In Record Settings, a Calibrate button listens to an external acoustic
calibrator (a 94 / 104 / 114 dB tone at 1 kHz) and records the microphone's
digital level at that known pressure; anything that is not a clean, dominant,
on-frequency tone is rejected rather than stored as a wrong number. What is
stored is the anchor's ingredients — the reference and measured levels, the
tone frequency, and the digital capture identity — not a baked "shift by N dB"
value: the Frequency Response is a loopback-referenced transfer function, so
turning the anchor into an SPL shift also uses each measurement's own loopback
level, while the Live Spectrum RTA needs only SPL = mic level + anchor offset.
The anchor is valid only at the gain it was captured at, so a changed digital input is flagged (the Calibrate button turns gold) and the dialog warns that the analog preamp gain must not move after calibrating.
Selecting dB SPL never depends on having an anchor, because the scale is also how you view curves captured in it: without one the plot keeps the dB SPL axis and becomes view-only — overlays recorded in dB SPL are drawn, the measurement's own curves are not (raw dBFS on an absolute axis would read as absurd pressures), and a notice on the plot says why. Starting a measurement in that state drops the display back to relative first, so a fresh run is never born hidden. The anchor is saved with the measurement settings and stamped onto every captured impulse response.
Resonalyze/
|-- source/ WinForms application: composition root, measurement
| | lifecycle, and plot presentation
| |-- History/ Measurement history snapshots and persistence
| |-- LiveSpectrum/ Live analyzer orchestration
| |-- Measurements/ Sweep/noise orchestration, signal generation, IR files
| |-- ModeSwitching/ The analysis-mode catalogue and tab controller
| |-- Options/ Measurement and visualization settings panels
| |-- Overlays/ Persistent overlay slots and calculated overlays
| |-- Plotting/ OxyPlot model creation, annotations, and adapters
| |-- Settings/ Settings file, schema migrations, update checking
| |-- Shell/ Main form, title bar, commands, and docked settings
| |-- TimeAlignment/ Loopback delay measurement UI and orchestration
| |-- Tools/ EQ Wizard, Signal Generator, Virtual DSP, PEQ import/export
| `-- Ui/ Reusable WinForms controls and dialogs
|-- dsp/ Reusable signal-processing library (no UI, no audio)
|-- audio/ Audio drivers and device access (NAudio lives here)
|-- tests/ App, audio, and synthetic DSP test projects
|-- installer/ Inno Setup script for the Windows installer
|-- assets/ Images used by the README and the application
|-- .github/workflows/ CI builds and automated tagged releases
|-- global.json Pinned .NET SDK version
`-- README.md
The three projects have deliberate boundaries. Resonalyze.Audio owns every
audio driver — MME, ASIO and both WASAPI modes — along with device enumeration,
format negotiation and capture lifecycle; NAudio is confined to it and is not
even referenceable from the application at compile time. Resonalyze.Dsp is pure
signal processing with no UI and no audio dependency: FFT analysis, windowing,
calibration, smoothing, impulse processing, phase analysis, group delay,
crossover and EQ design. The application project wires the two together.
- .NET 10, Windows Forms, OxyPlot
- NAudio and NAudio.Asio
- Math.NET Numerics
- NetSparkle — in-app updates
- YamlDotNet — CamillaDSP profiles
- PDFsharp / MigraDoc — tuning-sheet PDFs
Third-party package licenses are listed in THIRD-PARTY-NOTICES.md.
Bug reports, reproducible measurement cases, DSP corrections, and focused pull
requests are welcome. Known technical debt and improvement ideas are collected in
TODO.md — a good place to look for a first contribution. When
reporting a measurement issue, include the audio interface and driver, the sample
rate and bit depth, the measurement mode, the relevant analysis settings, the
expected and actual behavior, and a screenshot or exception stack trace —
unexpected errors are appended to crash.log in the application data directory.
Resonalyze is available under the MIT License.