Datasets:

---

uv-scripts

/

transcription

like 13

Follow

uv scripts for HF Jobs 171

Adds a new UV script that wraps kb-labb/easytranscriber to produce word-level timestamps via VAD → Cohere/Whisper ASR → wav2vec2 forced alignment.

Changes

• easytranscriber-transcribe.py — new script. Default backend is Cohere Transcribe 2B (same model as cohere-transcribe.py) but adds per-word timing. --backend {cohere,ct2,hf} switchable; --emit-txt / --emit-srt for side-files.
• README.md — new script row in Scripts table, per-script Options table, separate Benchmarks block.

Benchmark (L4, 66-file CBS Suspense, 33 h audio)

• 46.2 min wall, 42.9x realtime
• 42,633 segments / 295,468 words
• Cost ~$0.62

Notes

• Bundles static-ffmpeg in deps — HF Jobs base images lack the ffmpeg binary that easyaligner shells out to.
• Cohere backend requires transformers>=5.4.0 (pinned).
• Silero VAD is the default (no auth); --vad pyannote available for users who've accepted pyannote terms.

🤖 Generated with Claude Code