whisper-wrap 0.1.0

Lightweight faster-whisper batch transcription CLI for researchers

whisper-wrap

Lightweight Python CLI wrapper around faster-whisper for bulk transcription.

Features

• Batch transcribe audio/video files in a directory
• Output formats: txt, srt, vtt, json
• Always writes per-file output
• Optional merged output via --merge
• Metadata header included (file, duration if available, model, language, device, compute type)
• Progress + summary + per-file errors

Install (uv)

uv sync

Run without installing globally:

uv run whisper-wrap --help

Usage

uv run whisper-wrap /path/to/media \
  --model small \
  --language en \
  --device cpu \
  --compute-type int8 \
  --threads 8 \
  --output-dir outputs \
  --format txt \
  --recursive \
  --ext mp3 --ext wav --ext mp4

Publish

cp .env.example .env
# add PYPI_API_TOKEN to .env
./scripts/publish.sh

Harvard sentences

uv run whisper-wrap harvard --format text
uv run whisper-wrap harvard --format json --limit 10

Attribution: Harvard Sentences (1965 revised list) from https://www.cs.columbia.edu/~hgs/audio/harvard.html (IEEE Transactions on Audio and Electroacoustics, Vol. 17, Issue 3, 1969).

Common flags

• --model default: small
• --language optional; omitted means auto-detect
• --device one of cpu|cuda
• --compute-type one of float16|int8
• --threads CPU threads
• --output-dir output folder
• --merge also write a single merged.<format> file
• --format one of txt|srt|vtt|json
• --metadata/--no-metadata include metadata headers (default: on)
• --ext repeatable extension filter (with or without .)
• --recursive scan subdirectories

Notes

• faster-whisper runtime behavior depends on local FFmpeg and model availability.
• For cuda, ensure compatible CUDA runtime/drivers are installed.