meeting-transcribe

Runs whisper-stream (whisper.cpp) in the background, transcribing the chosen audio input in ~3-second steps into a growing text file under ~/Documents/MeetingTranscripts/. All management goes through one script:

scripts/transcribe.sh {check|setup|devices|start|status|stop}

Use the script rather than invoking whisper-stream directly — it handles the PID file, output naming, and fail-fast checks, and stop depends on its bookkeeping.

Workflow

1. Check readiness

scripts/transcribe.sh check — verifies whisper-stream and the model exist. If anything is missing, scripts/transcribe.sh setup installs whisper-cpp via Homebrew and downloads the multilingual base model (~148 MB). setup --model small fetches a more accurate model (~488 MB) — offer this if the user complains about transcription quality; then start with WHISPER_MODEL=~/.whisper-models/ggml-small.bin scripts/transcribe.sh start.

2. Pick the audio source

Default (no flags) captures the system-default input — the built-in mic. That is correct when the user talks in person or listens on speakers. If the user is in a call wearing headphones, the mic never hears the other participants: run scripts/transcribe.sh devices, and read references/system-audio.md to choose between the "Microsoft Teams Audio" virtual device and a BlackHole aggregate device. Ask the user which situation they're in if it isn't obvious.

3. Start

scripts/transcribe.sh start [--device N] [--lang auto|en|zh|...]

Language defaults to auto (fine for mixed Chinese/English meetings). The command prints the transcript path — tell the user where it is, and that they can watch it live with tail -f <path> in a terminal. First-ever run may pop a macOS microphone-permission dialog the user must accept.

If start FAILS, the script prints the engine log tail — the usual causes are a missing mic permission or an invalid device ID.

4. During the meeting

• scripts/transcribe.sh status shows the last transcribed lines — use it when the user asks "is it working?" or "what was just said?".
• To answer questions about the meeting so far, read the transcript file.
• Transcription runs independently of this conversation; nothing needs to stay open. Don't poll in a loop unless the user asks for live commentary.

5. Stop and summarize

scripts/transcribe.sh stop, then offer meeting notes. Read the transcript and produce:

# Meeting notes — <date>
## Summary        (3-6 bullets of what was discussed)
## Decisions      (each decision, with who/what/when if stated)
## Action items   (- [ ] owner: task — deadline if mentioned)
## Open questions

Write the notes next to the transcript as <transcript-name>-notes.md. Whisper output has no speaker labels and occasional mis-hearings — clean up obvious transcription errors from context, and don't invent owners for action items the transcript doesn't attribute.

Quality expectations

The streaming model favors latency over polish: expect duplicated fragments at chunk boundaries and weaker accuracy than offline transcription. For a high-stakes record, also pass --save-audio-style needs to a future run — or simply re-transcribe: if the user wants a cleaner final version and has an audio recording, run whisper-cli -m ~/.whisper-models/ggml-small.bin -f <audio> (convert to 16 kHz mono wav with ffmpeg first).