Voice Transcribe

Voice transcription (local, offline)

Transcribes audio to text on-device using whisper.cpp (small multilingual model) plus a pure-WASM Opus decoder. No network, no API key, works headless. Good for French and English.

Setup (one time)

bash ${CLAUDE_SKILL_DIR}/setup.sh

This installs the Node Opus decoder and builds whisper.cpp + downloads the small model (~466 MB) into the skill's engine/ folder. Idempotent — safe to re-run. Requires git, cmake, a C compiler, and node (18+).

When to use

• A message arrives with a voice attachment (audio/ogg, Telegram .oga).
• You have any audio file (.oga, .ogg, .wav, .mp3, .flac) and need its text.

How to use

${CLAUDE_SKILL_DIR}/bin/transcribe.sh <audio-path> [lang]

• lang is optional: fr, en, or auto (default). Force fr/en for best accuracy.
• The transcript is printed to stdout.

Example:

${CLAUDE_SKILL_DIR}/bin/transcribe.sh ~/Downloads/voice-note.oga fr

Notes

• First word of output may have a leading space — trim if needed.
• A ~15 s clip transcribes in a few seconds on CPU.
• To point at an existing whisper.cpp build instead of the bundled one, export VT_ENGINE=/path/to/whisper.cpp (must contain build/bin/whisper-cli + models/ggml-small.bin).
• For higher accuracy at the cost of speed/size, swap the model in setup.sh (e.g. medium).