transcribe

name	transcribe
description	Transcribe audio files locally using faster-whisper (no API key needed). Use this skill when you need to transcribe voice messages, audio files, or any speech-to-text task. Supports .ogg, .mp3, .wav, .m4a and other formats. Works offline.

Transcribe Skill

Local speech-to-text using faster-whisper. Runs on-device, no API key required.

Usage

uv run ${CLAUDE_SKILL_DIR}/transcribe_cli.py <audio_file>

Options

# Transcribe with auto language detection (default model: base)
transcribe_cli.py voice.ogg

# Use a larger model for better accuracy
transcribe_cli.py voice.ogg --model small
transcribe_cli.py voice.ogg --model medium

# Force a specific language
transcribe_cli.py voice.ogg --language en
transcribe_cli.py voice.ogg --language es

Models (accuracy vs speed tradeoff)

Model	Size	Notes
tiny	~75MB	Fastest, least accurate
base	~145MB	Good balance (default)
small	~460MB	Better accuracy
medium	~1.5GB	High accuracy
large-v3	~3GB	Best accuracy, slowest

Models are downloaded on first use and cached in ~/.cache/huggingface/.

Telegram voice messages

Voice messages from Telegram arrive as .oga files. Convert with ffmpeg first:

ffmpeg -i message.oga /tmp/voice.mp3 -y
uv run ${CLAUDE_SKILL_DIR}/transcribe_cli.py /tmp/voice.mp3