Transcription on this Mac
[stt] provider = "local" transcribes with whisper.cpp on the Mac: the audio never leaves it. The helper listens on 127.0.0.1 only and reads a model file from disk. Sidekick still needs an AI key for the coaching (cards, debriefs, reviews, the interview), unless you point it at a local server (below).
How it fits together
| Piece | File | What it does |
|---|---|---|
| Helper | backend/native/whisper/build.sh → bin/whisper-server | whisper.cpp's HTTP server, Metal, static, arm64, macOS 14.2+ |
| Supervisor | backend/hud/local_stt.py WhisperServer | starts the helper once with the chosen model, reuses it, stops it on quit |
| Live stream | LocalSTT | one decode per utterance (the VAD already cuts them); finals in order; whisper's silence phrases dropped |
| Batch | LocalClips via batch_stt | imported recordings; post-meeting re-transcription is skipped when live was already local |
| Models | Settings › Transcription, onboarding | downloaded on request into ~/Library/Application Support/MeetingHUD/models/ |
Building the helper
make whisperIt needs git and the Xcode Command Line Tools. cmake comes from your PATH or from PyPI through uvx. It downloads whisper.cpp at the tag in build.sh (WHISPER_TAG overrides it; check the project's releases before a release build). scripts/bundle-backend.sh copies the binary into the app (WHISPER=1 builds it first), and hud/scripts/sign-helpers.cjs signs it with the same identity as the audio helper. A build without it shows "On this Mac (not in this build)" and everything else works.
Models
| Key | File | Approx. size | Use |
|---|---|---|---|
large-v3-turbo-q5_0 | ggml-large-v3-turbo-q5_0.bin | 550 MB | default; multilingual (the bilingual case) |
small-q5_1 | ggml-small-q5_1.bin | 190 MB | faster, less accurate |
base-q5_1 | ggml-base-q5_1.bin | 60 MB | older Macs |
Downloads come from SIDEKICK_MODEL_BASE_URL (default: the whisper.cpp model repository on Hugging Face), are written to a .part file and renamed only when complete. The sizes above are approximate; the app shows the real size from the server.
Measurements (to do on real hardware)
Nothing below has been measured yet, and the README and the app make no speed or accuracy claims until it is.
| Mac | Model | Real-time factor | Card latency added | WER on the bilingual fixture |
|---|---|---|---|---|
| M1 8 GB | small-q5_1 | |||
| M1 8 GB | large-v3-turbo-q5_0 | |||
| M2/M3 16 GB+ | large-v3-turbo-q5_0 |
How to measure: hud doctor runs a self-test (a phrase synthesized to a file with say -o, never played aloud) and prints the decode time; for meeting-length numbers, run hud replay with FileAudioSource on recorded WAVs.
A local AI server for cards and reviews (experimental)
Settings › Transcription › Advanced takes an OpenAI-compatible address, for example Ollama's http://127.0.0.1:11434/v1 or LM Studio's. Sidekick uses the Responses API with structured output, so the server must support both, and [llm] openai_tick_model (cards) and openai_deep_model (everything else) must name its models. Small models often return answers the review cannot read; this is why it is not the default.