voice-transcription

name	voice-transcription
description	Transcribe voice messages and audio files to text using faster-whisper locally on CPU. Use when a user sends a voice note or audio file.

Voice Transcription

Transcribe audio files (voice messages, recordings) to text using faster-whisper with the turbo model running locally on CPU. No external API keys required.

Bundled Scripts

transcribe.py — Transcribe an audio file

# Basic transcription
uv run scripts/transcribe.py /path/to/voice.ogg

# Specify language (skip auto-detection)
uv run scripts/transcribe.py /path/to/audio.mp3 --language en

# Output as JSON (includes timestamps and language info)
uv run scripts/transcribe.py /path/to/audio.wav --json

Output (default):

Language: en (probability: 0.98)
[0.00s -> 3.50s] Hello, this is a test message.
[3.50s -> 6.20s] I'm recording this for transcription.

Output (JSON):

{
  "language": "en",
  "language_probability": 0.98,
  "segments": [
    {"start": 0.0, "end": 3.5, "text": "Hello, this is a test message."},
    {"start": 3.5, "end": 6.2, "text": "I'm recording this for transcription."}
  ],
  "full_text": "Hello, this is a test message. I'm recording this for transcription."
}

Setup

Dependencies (faster-whisper) are declared inline via PEP 723 and handled automatically by uv run. The turbo model (~800MB) is downloaded on first use and cached in ~/.cache/huggingface/.

Supported Formats

Any format supported by ffmpeg/PyAV: .ogg, .mp3, .wav, .m4a, .flac, .webm, .opus, etc.

Notes

• Uses turbo model (whisper-large-v3-turbo) — fast and accurate, optimized for CPU with INT8 quantization.
• First run is slower due to model download. Subsequent runs are much faster.
• Language is auto-detected but can be overridden with --language.