Private meeting summarization. Private ambient listening.
An iOS app that records meetings and produces structured notes entirely on-device. No cloud, no subscriptions — everything runs locally on your iPhone.
- Record a meeting using the built-in microphone
- Transcribe speech to text with Parakeet TDT v2 (~600MB CoreML model running on the Apple Neural Engine)
- Diarize speakers using Pyannote + WeSpeaker (~100MB CoreML models)
- Summarize the transcript into bullet-point notes with Gemma 4 E2B (3.1GB, Q4_0 quantized GGUF running via llama.cpp with full Metal GPU offload)
All models run on-device: ASR and diarization on the Neural Engine via CoreML, LLM inference on the GPU via Metal.
| Model | Task | Size | Runtime |
|---|---|---|---|
| Gemma 4 E2B Q4_0 | Summarization | 3.1 GB | llama.cpp (Metal GPU) |
| Parakeet TDT v2 | Speech-to-text | ~600 MB | FluidAudio (CoreML/ANE) |
| Pyannote + WeSpeaker | Speaker diarization | ~100 MB | FluidAudio (CoreML/ANE) |
A 13-minute recording (flue.wav) processed entirely on an iPhone 16 Pro.
Models are downloaded on demand from a Google Cloud Storage bucket and loaded into memory. Models persist across app updates.
Record a live meeting or upload a saved audio file.
After recording, the app automatically transcribes speech and identifies speakers.
Gemma generates structured bullet-point notes with clickable timestamp links that jump to the corresponding position in the transcript.
- iOS 17.0+
- iPhone with enough storage for ~4GB of models