Hugging Face bundles transcription scripts for GPU Jobs

The uv-scripts/transcription dataset on Hugging Face packages a growing collection of scripts that run audio-to-text models on GPU Jobs, pulling input from a Bucket or local folder and writing transcripts as .txt, JSON with speaker segments or word timestamps, .srt, or parquet. The dataset page describes them as "self-contained UV scripts that transcribe audio files on GPUs using Hugging Face Jobs."

Each script wraps a different model. `cohere-transcribe.py` runs the gated CohereLabs/cohere-transcribe-03-2026 across 14 languages. `buzzasr-transcribe.py` covers 102 FLEURS languages and can swap in `openai/whisper-large-v3` as a control. The heaviest, `nemotron-transcribe.py`, calls NVIDIA's Nemotron 3.5 Transcribe and offers, per the script notes, "Word timestamps and speaker tags from one model, 40 languages (speaker tags in 13)," with a fixed ~8-minute warm-up before the first file and an optional `--serve` flag for a live endpoint. MOSS-Transcribe-Diarize handles English and Chinese with speaker IDs under Apache 2.0.

The link was circulating among a couple of the researchers we track by the time we logged it. Running any of the scripts needs `hf` 1.32 or newer and pay-as-you-go Jobs credits on an L4, A100 or RTX-Pro-6000 flavor.