Transcription UV Scripts

Part of uv-scripts: self-contained UV scripts you run on Hugging Face Jobs in one command.

Scripts that transcribe audio files on a GPU with Hugging Face Jobs. Audio comes in from a Bucket or a local folder; transcripts go out as .txt, JSON (speaker segments or word timestamps), .srt or parquet.

Quick Start

First, install the hf CLI and sign in. Jobs needs pay-as-you-go credit. Use hf 1.32 or newer (hf version): older CLIs ignore the script's launch header without a warning, and the Job then runs on a CPU.

Transcribe four minutes of NASA's Houston We Have a Podcast (a host and a guest, from the demo Bucket) with Nemotron 3.5 Transcribe. Replace your-username with your Hugging Face username:

hf buckets create your-username/transcripts --private
hf jobs uv run --timeout 45m \
    -v hf://buckets/uv-scripts/transcription-demo/podcast:/input:ro \
    -v hf://buckets/your-username/transcripts/podcast:/output \
    https://huggingface.co/datasets/uv-scripts/transcription/raw/main/nemotron-transcribe.py \
    /input /output --timestamps

The Job writes a .txt with speaker tags and a .json with word timestamps to your Bucket:

<spk:0> Jim, thank you so much for coming on the podcast today. ... <spk:3> It is, uh, the mission control room, of course, is now open for business ...

Our test run took about 7.5 minutes on rtx-pro-6000 (about $0.34), almost all of it model setup. Read the result with hf buckets cp hf://buckets/your-username/transcripts/podcast/nasa-podcast-ep100.txt -.

More on this example: JSON output, MOSS on the same clip, harder audio, launch headers

JSON output. Every word has a start and end time and a speaker (parts[].alignment.words), plus diarization segments (parts[].alignment.diarization):

{"word": "sustainable", "start": 0.02, "end": 0.98, "speaker": "0", "part": 0}

MOSS on the same clip. MOSS-Transcribe-Diarize on the same clip took under 2 minutes on l4x1 (about $0.02) and wrote one line per segment ([61.82 - 67.06] S02: The mission control room, ...):

hf jobs uv run --flavor l4x1 -s HF_TOKEN --timeout 30m \
    -e UV_TORCH_BACKEND=cu128 \
    -v hf://buckets/uv-scripts/transcription-demo/podcast:/input:ro \
    -v hf://buckets/your-username/transcripts/podcast-moss:/output \
    https://huggingface.co/datasets/uv-scripts/transcription/raw/main/moss-transcribe-diarize.py \
    /input /output --emit-txt

Harder audio. Mount /apollo11 instead of /podcast: two clips of 1969 mission tapes with radio noise and crosstalk. The demo Bucket's README gives the sources and licences (NASA, public domain).

Why no flags. The Nemotron scripts carry a [tool.hf-jobs] header that sets the GPU, image and secrets. The other scripts have no header yet and take flags, as in the MOSS command. hf jobs uv run --dry-run <script> shows the resolved settings.

Use your own audio

Audio on your machine: mount the folder directly. -v ./audio:/input syncs it to a private Bucket and mounts it in the Job; re-runs only sync changed files (huggingface_hub ≥ 1.22). Start with --max-files 2.

Audio in a Bucket: replace the demo input with -v hf://buckets/your-username/your-audio:/input:ro. The scripts walk the input folder recursively and mirror its layout in the output (easytranscriber puts its JSON under alignments/). nemotron-transcribe-saturate.py is the exception: it takes a glob ("hf://buckets/user/audio/**/*.mp3") and writes parquet.

Audio on the Internet Archive: download-ia.py downloads an item's audio straight into a mounted Bucket:

hf jobs uv run \
    -v hf://buckets/your-username/audio-files:/output \
    https://huggingface.co/datasets/uv-scripts/transcription/raw/main/download-ia.py \
    SUSPENSE /output

Local dev: if you've cloned this repo, swap the URL for the local filename (e.g. cohere-transcribe.py /input /output ...).

Pick a model

You need	Script	Trade-off
Punctuated text in one of 14 languages (en, de, fr, it, es, pt, el, nl, pl, ar, vi, zh, ja, ko)	cohere-transcribe.py	Gated: accept the model terms first
A lot of English audio, as fast as possible	granite-turboctc-transcribe.py	Lowercase, no punctuation
A language Cohere doesn't cover (102 FLEURS languages)	buzzasr-transcribe.py	One model per language; mostly lowercase, no punctuation
Word-level timestamps (subtitles, search, alignment)	easytranscriber-transcribe.py	Four-stage pipeline, slower
Who said what, in English or Chinese	moss-transcribe-diarize.py (-server.py for many files)	Slow on long files; server version needs a100-large
Word timestamps and speaker tags from one model, in 40 languages (speaker tags in 13)	nemotron-transcribe.py	~8 min fixed setup before the first file
Nemotron text for a large batch, resumable	nemotron-transcribe-saturate.py	Plain text only

Only the two Nemotron scripts set their own GPU and image. The others need flags, for example Cohere on an L4:

Cohere Transcribe command

Accept the terms on the model page first; the model is gated.

hf jobs uv run --flavor l4x1 -s HF_TOKEN \
    -e UV_TORCH_BACKEND=cu128 \
    -v hf://buckets/your-username/audio-files:/input:ro \
    -v hf://buckets/your-username/transcripts:/output \
    https://huggingface.co/datasets/uv-scripts/transcription/raw/main/cohere-transcribe.py \
    /input /output --language en --compile

Each script's docstring has its own hf jobs uv run example.

MODELS.md has the details for each script: speed, quality notes, pins and known issues. BENCHMARKS.md has the measured runs.

Serve as a live endpoint

Nemotron and MOSS-Transcribe-Diarize can run as a server in a Job with an exposed port, so you pay the setup once and send requests from a laptop or an agent. Start one with nemotron-transcribe.py --serve; SERVING.md has the commands, request examples and caveats.

Options

The folder scripts take /input /output and --max-files N for a test run (the saturate script takes a glob and an output dataset). --help lists all flags.

nemotron-transcribe.py

Flag	Default	Description
--style	verbatim	Model-card prompt preset: verbatim (fillers kept), clean (fillers dropped) or speakers (<spk:N> tags)
--prompt	none	Custom user prompt (overrides --style), e.g. Transcribe with punctuation and capitalization. Context terms: <names>.
--timestamps	off	Also write JSON with word timestamps, speaker tags and diarization segments (uses the speakers prompt)
--concurrency	16	Files in flight against the server
--part-minutes	55	Split longer audio into equal parts of at most this length, up to 60 (the model is trained on recordings up to 1 h); speaker tags reset between parts
--server	http://127.0.0.1:8000	Use a server already running there instead of starting one; a serving job's https://<job_id>--8000.hf.jobs gets your HF token as Bearer auth (no other host does)
--serve	off	Only start the server and keep it up, for a job started with --expose 8000 (no input/output dirs)
--nemo-ref	the image's commit (7b05f096)	NeMo Speech ref to install when the image has no prebuilt stack; main follows the model card
--workdir	/opt/nemotron-transcribe (prebuilt image), else /tmp/nemotron-transcribe	Where the serving stack and checkpoint live (the prebuilt image's stack is used when present)
--request-timeout	3600	Per-request timeout in seconds
--max-files	all	Limit files to process (for testing)

cohere-transcribe.py

Flag	Default	Description
--language	required	en, de, fr, it, es, pt, el, nl, pl, ar, vi, zh, ja, ko
--compile	off	torch.compile encoder (one-time warmup, faster after)
--batch-size	16	Batch size for inference
--max-files	all	Limit files to process (for testing)

cohere-transcribe-vllm.py

Flag	Default	Description
--language	required	Same 14 languages as cohere-transcribe.py
--gpu-memory-utilization	0.8	Fraction of GPU memory vLLM may use
--verbose	off	More logging
--max-files	all	Limit files to process (for testing)

granite-turboctc-transcribe.py

Flag	Default	Description
--chunk-seconds	0 (whole file)	Cut files into N-second windows; use 30 for hour-plus files or 24 GB cards (2.2 GB peak at batch 32 vs 10.3 GB for four 30-min files)
--batch-size	4	Files (or windows) per forward pass; 32-64 with --chunk-seconds 30
--decode-workers	CPU count	Threads for mp3/wav decoding, overlapped with model load and inference
--max-files	all	Limit files to process (for testing)

buzzasr-transcribe.py

Flag	Default	Description
--language	required	Language name (dutch, sorani-kurdish) or ISO code (nl, ckb), resolved via the org's model tags; --list-languages prints all of them live from the Hub
--model	BuzzASR/<language>	Any Whisper checkpoint on the Hub (e.g. openai/whisper-large-v3 as a zero-shot control)
--long-form	sequential	sequential (Whisper's own long-form decoding with timestamps + temperature fallback; skipped no windows in our tests) or chunked (transformers pipeline, 30 s windows with 5 s overlap; ~2.5x faster, drops text at bad merges)
--batch-size	16	Files per generate call (sequential) or 30 s windows per forward pass (chunked)
--decode-workers	CPU count	Threads for mp3/wav decoding, overlapped with model load
--max-files	all	Limit files to process (for testing)

easytranscriber-transcribe.py

Flag	Default	Description
--language	required	ISO 639-1 code. Cohere supports the same 14 languages as cohere-transcribe.py; ct2/hf support any Whisper language
--backend	cohere	cohere, ct2 (CTranslate2 Whisper, fastest for Whisper), or hf (transformers)
--transcription-model	Cohere 2B / distil-whisper-large-v3.5	HF model ID; override to use KB-Whisper, Whisper-large-v3, etc.
--emissions-model	per-language default	wav2vec2 for forced alignment: en→wav2vec2-base-960h, sv→voxrex-swedish, else→facebook/mms-1b-all
--vad	silero	silero (no auth) or pyannote (requires accepting terms + HF_TOKEN)
--tokenizer-lang	derived from --language	NLTK Punkt language name for sentence tokenization
--emit-txt	off	Also write .txt transcripts alongside the JSON alignments
--emit-srt	off	Also write .srt subtitles derived from alignment segments
--batch-size-features	8	Feature-extraction batch size
--batch-size-transcribe	16	ASR batch size (where backend supports it)
--max-files	all	Limit files to process (for testing)

moss-transcribe-diarize-server.py

Flag	Default	Description
--concurrency	4	Concurrent transcription requests (KV-cache bound; 6 works on A100)
--part-minutes	55	Clip length for splitting long audio (longest that fits the context window). Speaker labels reset between clips
--server	http://127.0.0.1:8000	sgl-omni server URL (in-job localhost by default)
--request-timeout	3600	Per-request timeout in seconds
--emit-txt	off	Also write .txt transcripts
--max-files	all	Limit files to process (for testing)

moss-transcribe-diarize.py

Flag	Default	Description
--max-new-tokens	0 (auto)	Token budget per file. Auto scales with audio duration (min 5120, max 65536). Output JSON sets "truncated": true if the budget was hit — re-run with a higher value
--hotwords	none	Comma-separated terms (names, companies, jargon) appended to the prompt to bias recognition
--prompt	built-in	Full prompt override (replaces the built-in transcribe+diarize prompt)
--emit-txt	off	Also write .txt transcripts ([start - end] SPEAKER: text per line)
--emit-srt	off	Also write .srt subtitles with speaker prefixes
--max-files	all	Limit files to process (for testing)

Notes

• Gated model: Accept terms at the model page before use.
• easytranscriber: the Cohere backend requires transformers>=5.4.0 (pinned in the script). Pyannote VAD is gated — accept terms at pyannote/segmentation-3.0 and pyannote/speaker-diarization-3.1 if using --vad pyannote. Otherwise stick with the default Silero VAD.
• MOSS-Transcribe-Diarize: not gated (Apache 2.0). Runs remote code from the model repo; see MODELS.md.