wav2sum

Audio → speaker-attributed transcript → meeting summary. Runs locally (MPS / CUDA / CPU).

Install

uv sync
uv tool install -e .
ollama pull gemma4:26b-a4b-it-q4_K_M
export HF_TOKEN="hf_..."

Default summary model: gemma4:26b-a4b-it-q4_K_M (override with --model).

Usage

wav2sum meeting.wav                                          # speakers default to 0=Я, 1=Илья
wav2sum meeting.wav --speaker "0=Alex" --speaker "1=Maria"   # override speaker names
wav2sum meeting.wav --model llama3.1:8b --no-timestamps -v

# Speaker count (default: 2). A range turns on auto-detection.
wav2sum meeting.wav --num-speakers 4
wav2sum meeting.wav --min-speakers 2 --max-speakers 5

# Reuse cached diarization/transcription off; recompute everything.
wav2sum meeting.wav --no-cache

Writes transcript.txt, summary.md and meta.json to output/.

Record a call (wav2sum record)

Records your mic + the other side's system audio into one wav. Stop with Ctrl-C.

wav2sum record                       # → ./recordings/call-<timestamp>.wav
wav2sum record --summarize           # record, then transcribe + summarize
wav2sum record -o zoom.wav --summarize

Device names are configurable via --output-device / --input-device / --headphones.

One-time macOS setup (Audio MIDI Setup):

• brew install blackhole-2ch switchaudio-osx
• Multi-Output "Call output" = headphones + BlackHole 2ch (headphones as Master, Drift Correction on BlackHole). Select it as the speaker in your call app.
• Aggregate "Call recorder" = built-in mic + BlackHole 2ch.