Voice, disciplined.
Local-first dictation and meeting assistance for Windows. Turn speech into text, keep searchable notes, and use optional AI formatting without requiring a paid AI service.
Build from source · Features · Privacy
The Meetings workspace. Screenshots use synthetic demo data.
- Dictation: resident Vosk recognition for Instant mode, or Whisper with optional Vulkan acceleration for Accurate mode. Transcription begins while speaking, using a bounded rolling audio buffer.
- Formatting: deterministic cleanup and an optional local Ollama model, with configurable formatting strength, a personal lexicon, and app profiles.
- Meetings: computer audio and microphone together, with source labels and timestamps; optional local translation preserves the original transcript. A resizable companion offers transcript, assistant, and split views. Asking the assistant includes newly captured transcript context without stopping capture. Optional automatic commentary runs only through a verified local Ollama model.
- Investigations: long-lived collections of pasted text, notes, screenshots, links, voice notes, and pinned answers, shown as a chronology or a connected diagram. Screenshot text is read by locally downloaded OCR models. A dedicated shortcut sends the clipboard to the active investigation, and the assistant answers with card citations.
- Library: exact-word and optional local semantic search over retained text.
- Actions: saved HTTP requests with transcript substitution or optional model-generated JSON, accessible from the launcher.
- Desktop controls: configurable shortcuts, light/dark appearance, model management, and setup/repair tools.
- Operations: review-and-send voice feedback and opt-in signed update support. These require configured server storage and release signing before public use.
The normal, resizable companion puts the live transcript and assistant beside the conversation. Choose Original, Translation, or Both; resize the split or focus on one pane. Ask a question with the latest transcript context while capture continues. Independent controls pause Phorminx's transcription of your mic or computer audio; they do not mute your call.
Phorminx targets Windows 10/11 x64. It is alpha software; recognition accuracy, latency, and compatibility depend on the device, audio source, and model.
Build from source using the development guide. Speech models and Vosk runtime assets are separate downloads and are not included in this source repository. Configure them in the application's setup and model screens. Ollama is optional; dictation can run without AI formatting.
The branded installer opens Phorminx's setup window after installation. Choose Prepare everything to install the local starter bundle without a terminal, or review individual repair steps. Progress and errors appear in Phorminx; a safe stop finishes the current step first. Use Prepare Phorminx in the Start menu to retry. Silent updates skip this onboarding. Setup preserves existing settings and saved text.
For an already installed app, an English starter bundle can be downloaded and configured automatically. Close Phorminx (including its tray icon), then run this from the repository folder in 64-bit PowerShell as your normal user:
powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\scripts\Initialize-Phorminx.ps1The script downloads Vosk's English model/runtime, Whisper base.en, installs
Ollama and the Vulkan runtime if missing, and pulls qwen2.5:1.5b for formatting
and nomic-embed-text for semantic search. Allow several GB of free disk space
and time for downloads. Windows may request administrator approval for Vulkan.
The speech assets are hash-verified and installers are publisher-verified.
On a fresh installation it configures Instant mode, balanced formatting, and
local semantic search. Existing settings and saved text are preserved; existing
installations may need to select the downloaded models in Settings/Models.
The script can be rerun after a failed download. Add -Plan to preview its scope
without changing anything. It is also standalone: the script file can be shared
alongside the app installer, without requiring a repository checkout.
This does not install Phorminx itself, source-build tools, or GPU drivers. Vulkan acceleration needs a compatible GPU/vendor driver; installing the Vulkan runtime alone does not establish GPU compatibility. This starter downloads English models, not every available language or model variant.
The default Ctrl+Alt+Space shortcut opens the P launcher. Choose Dictate, speak,
then press the shortcut again to finish. Settings provides shortcut and
appearance preferences. Meetings and Actions have their own workspaces.
Models also offers curated English Vosk choices ranging from a small download to multi-gigabyte models. Downloads do not switch the active model: explicitly choose Use for English Instant after setup. Native validation runs in the background. Larger models require more memory and are not guaranteed to improve every voice. Dual-source meetings run two recognizers; headphones help avoid speaker audio being transcribed a second time through the microphone. Meeting settings save the input and output devices separately for the next session. A configured, installed small local fallback model may retry a failed assistant response once; assistant failure never ends capture.
Speech recognition runs locally. The production capture path keeps audio in
memory rather than saving recordings to disk. Retained text and settings live
under %LOCALAPPDATA%\Phorminx; use the application's history, session, and
deletion controls to manage stored text. Disabling dictation history does not
mean that separately saved meeting sessions have been deleted.
Local Ollama formatting and embeddings use local compute. Model downloads use the network. Optional remote assistant providers and HTTP actions send the configured text/context to the selected provider or destination. These options are not a local-only privacy guarantee: remote providers may charge for usage and apply their own data policies. Review the destination before enabling them.
Meeting translation uses a separately selected local Ollama model. Originals
remain available even if translation is unavailable or cannot keep up. Fresh
settings recommend nomic-embed-text:latest; existing embedding selections and
indexes are not silently replaced.
Investigations are kept until you delete them; dictation history retention does not apply to them. Screenshots are stored as re-encoded PNG files beside the database. Screenshot text recognition runs locally after an optional, verified model download (12.6 MB). A pasted link is fetched once, without cookies, for its title and preview; local and private-network addresses are never fetched.
Messages (off by default) sends text directly to people you add with a one-time invite. Turning it on connects Phorminx to the Tor network through the embedded Arti client; there is no Phorminx server, and neither side learns the other's IP address. Contacts can see when you are online, and a message only moves while both of you have Phorminx running. History stays on this computer and can be deleted per contact. See ADR-018.
Launcher 0 · Give feedback records a draft locally. Only clicking Send
submits its reviewed text to phorminx.net; optional diagnostics are shown
separately. Audio and logs are not attached automatically. Release-enabled
builds check stable GitHub Releases daily by default; automatic download and
install after exit are on by default, with an onboarding opt-out and a Settings
control. The app verifies signed metadata and installer hashes and
waits for exit before installing, never stopping an active recording. Builds
without a configured release-verification key cannot enable updates. Neither
feature makes an unconfigured feedback server ready.
Do not include real transcripts, recordings, credentials, or personal settings in bug reports or contributions. Use synthetic examples instead.
- Build, test, and package
- Windows compatibility checks
- Architecture decision records
- Production brand assets
- Assistant transport contracts
- Meeting capture and translation
- Setup and operations release gates
- Setup and meetings acceptance checklist
- Investigations and local OCR
Phorminx's original source is licensed under the MIT License. Third-party libraries, runtimes, and downloaded models retain their respective licenses and notices; the project license does not replace those terms.