Witness

Witness is a local meeting recorder and transcriber for Windows 11. It can detect Microsoft Teams calls, capture microphone and system output on separate tracks, transcribe on the device, identify recurring speakers, and keep a searchable archive.

Meeting audio, transcripts, notes, and voice prints are not sent to a Witness service. Choosing Download for a speech model connects to Hugging Face; choosing Download GPU libraries connects to PyPI's files.pythonhosted.org host for pinned NVIDIA-published wheels. No meeting content is included in those requests. See Privacy before recording other people.

Witness is pre-1.0 software. Test it with non-critical recordings and keep independent backups of anything you cannot replace.

What it does

• Automatically starts and stops around Teams calls, or records manually.
• Captures the selected microphone and render-device loopback as two tracks. Loopback includes every sound sent to that output device, including music and notifications. Route Teams to a dedicated output device if isolation matters.
• Produces a stereo Ogg Opus archive, normally around 18 MB per hour at the default bitrate.
• Runs Parakeet TDT or Whisper transcription locally. Parakeet supports live captions; the final transcript replaces those provisional captions.
• Diarizes up to four remote speakers and stores explicit speaker enrollments as local voice-print embeddings.
• Searches transcript text and notes with SQLite FTS5.
• Recovers interrupted recordings and queued processing after a restart.
• Moves deleted meetings to a recycle bin before permanent deletion.
• Creates atomic, checksummed recovery snapshots from Settings.

Install and first run

Witness currently targets Windows 11. Install a release build, start Witness, and leave its tray process running. Closing the main window hides it; use Quit from the tray to exit.

Before an important call:

1. Open Settings > Recording and select the intended microphone and output device. Saved endpoint IDs are strict: if a configured device is missing, Witness reports it instead of silently recording another device.
2. Make a short manual test recording and play both sides back.
3. Download the desired models under Settings > Models.
4. Confirm local rules and participant consent allow recording.

Recording works without models; transcription, diarization, voice matching, and live captions require their corresponding downloads. A tray notification announces automatic recording. The top bar and Diagnostics show capture loss or device failures.

Models, network access, and disk space

Model downloads are explicit and pinned to immutable upstream revisions. Each file is size- and checksum-verified before it is considered installed. A failed or interrupted download remains incomplete and can be repaired from Settings.

Allow at least 5 GB free when installing every model so downloads and temporary files have headroom. Parakeet attempts NVIDIA CUDA and falls back to CPU if the provider cannot initialize. GPU transcription needs the CUDA 13 runtime and cuDNN 9 resolvable by the DLL loader; Witness checks this before every engine load, automatically picks up installs referenced by CUDNN_PATH or placed in the standard NVIDIA\CUDNN layout (under Program Files or LOCALAPPDATA), and otherwise logs which DLLs are missing. Settings shows the same GPU status.

Machines with an NVIDIA GPU but no CUDA developer install can use Settings → Models → Download GPU libraries (~1 GB): Witness fetches the pinned, SHA-256-verified NVIDIA-published CUDA runtime, cuDNN, and nvJitLink wheels hosted by PyPI (files.pythonhosted.org) and installs just the required DLLs into the data directory — no admin rights, no system changes, removable by deleting cuda/. GPU transcription can use the new libraries immediately, without a restart.

Whisper is CPU-only in the default build. Model revision and integrity status appears in Settings and the privacy-safe diagnostic report.

Storage and retention

witness-settings.toml is read from WITNESS_SETTINGS (a file, or a directory containing the default file) when set, otherwise from beside witness.exe. The configured data directory contains:

witness.db       transcripts, notes, queue state, and voice prints
audio/           archived .opus recordings
rec-tmp/         in-flight/recoverable WAV files and sidecars
models/          downloaded model cache and integrity metadata
cuda/            optional app-managed CUDA/cuDNN/nvJitLink GPU libraries
witness.log      rotating operational log

Changing the data directory applies after restart and does not move existing data. Deleted meetings remain in the recycle bin for 30 days, then are purged at startup. Emptying the bin is immediate and irreversible. Backups are independent copies and are not affected by retention actions in the app.

The database, audio, settings, logs, and backups are not encrypted by Witness. Use Windows device encryption/BitLocker and an account with a strong sign-in if the archive is sensitive.

Backups and recovery

Use Settings > Storage > Create backup while recording and processing are idle. Witness writes into a private staging directory and publishes a Witness-backup-... directory only after every payload and its checksum manifest is complete. The snapshot includes:

• a consistent standalone witness.db snapshot, including transcripts, notes, recycle-bin state, queue state, and voice prints;
• archived audio and recoverable rec-tmp files;
• the persisted settings; and
• model revision/integrity metadata.

Large model binaries are intentionally excluded and can be re-downloaded. Never choose a destination inside the active Witness data directory. Full verification and non-destructive restore instructions are in Backup and recovery.

Diagnostics and troubleshooting

Open Settings > Diagnostics to inspect the active data/database/log paths, SQLite health, capture loss, processing stage, queue depth, and model integrity. The copied/exported support report contains operational state and local paths, but no transcript text, notes, audio, voice prints, or private settings. Review paths before sharing it.

Common problems:

• Configured audio device missing: reconnect it or deliberately choose a new endpoint in Settings. Witness will not silently fall back.
• Output track is silent: confirm Teams is routed to the output device that Witness captures. Per-process loopback is not used because it is unreliable for Teams.
• Unrelated audio was recorded: loopback captures the complete selected output device. Route Teams to a dedicated virtual or hardware channel.
• Recording shows dropped time/device losses: preserve the recording, export Diagnostics, and check USB/Bluetooth power, drivers, and device changes.
• Transcription is unexpectedly slow: check the GPU line in Settings and witness.log — a warning names any CUDA/cuDNN DLLs the loader cannot find, and a single-digit realtime factor in completion toasts means CPU. Use Download GPU libraries in Settings → Models (or install cuDNN 9 for CUDA 13 / set CUDNN_PATH) to restore GPU transcription. CPU fallback is functional but slower.
• Model integrity check failed: use Repair in Settings. Witness will not load a model until the pinned files verify.
• Meeting remains failed: keep its rec-tmp files and use Retranscribe. Do not purge the meeting until recovery succeeds.
• New data directory appears empty: the selection does not migrate data. Return to the old directory or follow the recovery guide.

Build from source

Requirements:

• Rust stable with the MSVC target;
• Node.js 24.15.x and npm (the supported range is >=24.15 <25);
• Visual Studio Build Tools with Desktop C++ and CMake; and
• LLVM/libclang 19.x for bindgen. Put CMake on PATH and set LIBCLANG_PATH to your LLVM bin directory when it is not discoverable. Do not rely on machine-specific paths in repository configuration.

npm ci
npm run build
Set-Location src-tauri
cargo test --locked
cargo clippy --locked --all-targets -- -D warnings
Set-Location ..
npm run tauri dev

Build the NSIS installer with npm run release:windows. The command compiles first, verifies the generated ONNX Runtime provider DLLs, and only then bundles them through the release configuration. The base Tauri configuration keeps bundling disabled so a one-phase tauri build cannot silently omit those DLLs. The default Whisper build uses CPU; --features whisper-cuda requires a compatible CUDA toolkit. Runtime CUDA support for Parakeet also needs a compatible NVIDIA driver and cuDNN 9.

See Contributing and the release checklist before proposing a change or shipping an artifact.

Limitations

• Windows 11 only; Teams is the only built-in auto-detection target.
• No echo cancellation. Speakers can bleed remote audio into the microphone track; a headset is recommended.
• Diarization supports at most four remote speakers.
• Voice matching is probabilistic. Approximate matches are marked and should be confirmed or corrected.
• Live captions are provisional and are not backfilled after a UI reload.
• No application-level encryption or cloud synchronization.

Security and license

Report vulnerabilities as described in SECURITY.md. Witness is released under the MIT License.