Local audio transcription CLI using faster-whisper. Runs entirely offline on your machine.

Requires Python 3.12+ and uv.

cd ~/Developer/whisper-tool
uv sync

# Plain text output
uv run transcribe recording.m4a

# JSON with segment timestamps
uv run transcribe recording.m4a --json

# Word-level timestamps (implies --json)
uv run transcribe recording.m4a --words

# Batch transcription to a directory
uv run transcribe *.m4a --output-dir ./transcripts/

# Use a larger model for better accuracy
uv run transcribe recording.m4a --model large-v3

Plain text (default) prints the transcription to stdout. When multiple files are given, each is preceded by a --- filename --- divider.

JSON (--json) returns structured output:

{
  "file": "recording.m4a",
  "language": "en",
  "language_probability": 0.98,
  "duration_seconds": 62.4,
  "text": "Full transcribed text...",
  "segments": [
    { "start": 0.0, "end": 4.8, "text": "First segment..." }
  ],
  "stats": {
    "segment_count": 12,
    "word_count": null,
    "transcription_ms": 3200
  }
}

With --words, each segment includes a words array containing per-word start/end times and confidence probabilities. Multiple files produce a JSON array.

Larger models are slower but more accurate. The base default is a good starting point. Bump to small or medium for noisy audio or accented speech. large-v3 gives the best quality at significant compute cost.