Your meetings, transcribed entirely on your Mac.
Captures your mic and the far side of the call (Zoom, Teams, Meet), and turns them into speaker-attributed transcripts that never leave your machine. 100% open-source, open-model, airgapped.
Why it exists
Cloud meeting AI means uploading the raw audio of every conversation to someone else's servers. For anything confidential, that's a non-starter.
So I built the opposite: transcription that runs entirely on-device, on open models, with nothing phoning home. And the transcripts aren't just notes — they're private context my own AI agents can draw on for total recall, without renting my memory to anyone. parley covers calls and meetings; mailrag covers email. Independent tools; my agents know about both and use what fits.
What it does
Dual-stream capture
Records your mic and system audio as separate streams, so local and remote voices stay distinguishable. No virtual drivers.
On-device diarization
Automatic who-said-what (pyannote + WeSpeaker + VBx), with a quality score per segment.
Two engines
FluidAudio / Parakeet (fastest, 25 EU languages) or Apple SpeechAnalyzer. Swap in Settings.
Echo / mic-bleed removal
Strips the far-end voice that bleeds into your mic so it isn't mistaken for a phantom speaker.
Crash-safe recording
Survives UI and XPC crashes with auto-relaunch, silent re-attach, and multi-segment stitching.
Open formats
JSON, SRT, and TXT with timestamps, speaker labels, confidence scores, and local/remote tags.
How it works — the hard parts
- Dual-stream capture (the core constraint). macOS exposes no API for a pre-mixed mic + system stream — verified through the macOS 26 SDK. So the app runs two ScreenCaptureKit streams and treats "local" and "remote" as first-class. That's what makes reliable speaker separation possible.
- Echo / mic-bleed removal. A triple-confirmed gate removes the far-end voice bleeding into your mic: >50% temporal overlap and >70% word overlap and >0.8 speaker-embedding cosine.
- Cross-chunk speaker reconciliation. Audio is chunked and transcribed in parallel; per-chunk speaker IDs are merged into one global identity via greedy cosine matching on embeddings.
- Crash-safe by design. Sentinel file + LaunchAgent restart + multi-segment stitching mean a crash mid-meeting costs ~300–800 ms, not your recording.
- Fully on-device ML. Parakeet ASR + pyannote/WeSpeaker/VBx diarization + VAD, all on the Apple Neural Engine.
Open models, on-device
| Component | Model | License |
|---|---|---|
| Speech recognition | NVIDIA Parakeet TDT 0.6B (CoreML) | CC-BY-4.0 |
| Speaker diarization | pyannote segmentation + WeSpeaker | CC-BY-4.0 |
| Voice activity | Silero VAD | MIT |
| Engine SDK | FluidAudio | Apache-2.0 |
Get started: git clone https://github.com/fmasi/parley.git && cd parley && bash package_app.sh --install