be9f14ce34
A small helpdesk system: an office Pixel running a patched GrapheneOS Dialer answers technician calls, records both call legs as separate channels, and a Ruby backend transcribes them through Whisper and files an AI summary against the caller. Squashed to a single commit for sharing. No credentials are included; secrets live outside the repo in /etc/helpdesk/env on the server or a gitignored .claude/env.local locally. See .claude/env.local.example for the shape. Start at README.md, then docs/architecture.md. |
||
|---|---|---|
| .. | ||
| README.md | ||
| requirements.txt | ||
| smoke_test.sh | ||
| transcribe.py | ||
| watch_progress.sh | ||
transcribe.py - Czech audio → text with speaker tags
Offline Czech speech-to-text. Output is plain [Speaker N]: … turns.
Built on WhisperX (Whisper large-v2 + Czech wav2vec2 alignment + pyannote
speaker-diarization-community-1). 100% local - no cloud calls.
One-time setup (~5 min + ~5 GB model download on first run)
sudo apt install ffmpeg # Debian/Ubuntu; use your distro's pkg mgr otherwise
cd whisper
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
The Hugging Face token in .env is mine and the gated-repo terms are
already accepted on that account, so it just works. No HF account or
clicks needed on your side. If you'd rather use your own token, replace
HF_TOKEN in .env, then visit (logged in) and click "Agree" on:
- https://huggingface.co/pyannote/speaker-diarization-community-1
- https://huggingface.co/pyannote/segmentation-3.0
Run
set -a; source .env; set +a # load HF_TOKEN
source .venv/bin/activate # if not already active
python transcribe.py recording.m4a # → recording.txt
Common flags:
--timestamps prepend [HH:MM:SS] to each turn
--min-speakers 2 lower bound (default 2)
--max-speakers 3 upper bound (default 3)
--model large-v3 alternative model (default: large-v2)
--language en non-Czech audio
-o out.txt custom output path
Accepts any format ffmpeg can read: mp3, wav, m4a, ogg, flac, mp4, …
Speed
- NVIDIA GPU: ~realtime (10 min audio → ~10 min processing)
- CPU only: 5-20× slower (10 min audio → 1-3 hours). The tool warns and proceeds automatically.
Live progress bar (optional)
In a second terminal while a run is going:
bash watch_progress.sh
Smoke test (optional)
bash smoke_test.sh - needs espeak-ng (sudo apt install espeak-ng).
Synthesizes a fake 2-speaker Czech dialogue and runs the full pipeline.
TTS audio so the transcription is robotic gibberish; only verifies plumbing.
Exit codes
0 ok · 2 bad args / missing token · 3 no ffmpeg · 5 HF terms not
accepted · 6 alignment model failed · 7 diarization failed · 130 Ctrl-C