babelscribe

Transcribe any audio or video, in any of Whisper's 99 languages, on any GPU — AMD, NVIDIA, Intel or Apple — or just the CPU. One command, no CUDA required, subtitles out.

pip install babelscribe          # Python 3.9+
babelscribe interview.mp4                 # auto-detects language, picks your best GPU -> interview.srt + interview.json

babelscribe is a thin, friendly layer on top of whisper.cpp. It exists because getting fast Whisper on a non-NVIDIA card (e.g. an AMD Radeon on Windows) still means building whisper.cpp with the Vulkan SDK yourself. babelscribe downloads a prebuilt whisper-cli for your system and handles the rest.

A 19-minute English TED talk transcribes in 48 seconds on an AMD Radeon RX 9070 XT with 1.5% word error — see the benchmark below.

Benchmark: real talks, human captions as the answer key

Four TED / TEDx talks, scored against the human-made captions in the spoken language (bench/bench.py, reproducible). Error = word error rate (WER) for space-separated languages, character error rate (CER) for Japanese and Thai. GPU: AMD Radeon RX 9070 XT via Vulkan. default = large-v3-turbo; accurate = --accurate (see the FLEURS section).

How to read it: TED captions are edited for reading (fillers dropped, light rewording), so these numbers are an upper bound — most of the Spanish "errors" are the speaker's actual words versus the tidied caption. Thai is genuinely harder: fast, casual speech with slang; --accurate (Pathumma Whisper text + turbo timing) cuts its error by more than a quarter. Talks are used only to measure accuracy; their transcripts are not redistributed (TED content is CC BY-NC-ND).

Benchmark: 16 languages on FLEURS

Google FLEURS test set, 50 utterances per language, human-verified verbatim transcripts (bench/fleurs.py, reproducible). Same normaliser for every language (Whisper's rule: lower-case, drop punctuation and non-spacing marks, numbers spelled out). WER for space-separated languages, CER for ja / zh / ko / th.

* Some FLEURS clips contain more speech than their reference transcript, so a correct model is charged for words the reference leaves out. Edge-trimmed ignores extra words before the first / after the last reference word; both numbers are stored by bench/fleurs.py. 50 utterances per language means differences under ~0.5 points are noise.

Also measured and not used, because large-v3 was as good or better: large-v2 (all languages), PhoWhisper-large (vi 19.0%), whisper-large-v3 dialectal / code-switching Arabic fine-tunes (12.2% / 15.2%), Typhoon Whisper (th 11.9%), Thonburian Whisper (th 9.1% — kept as an option), Vaani Hindi (16.5%).

Why babelscribe (vs. what already exists)

babelscribe does not replace those projects — it stands on whisper.cpp and simply removes the hard parts: compiling for your GPU, converting media, picking the right device, avoiding the long-file repeat bug, and combining a language-specific fine-tune with accurate timestamps.

Languages

All 99 languages Whisper was trained on, auto-detected or forced with -l: af am ar as az ba be bg bn bo br bs ca cs cy da de el en es et eu fa fi fo fr gl gu ha haw he hi hr ht hu hy id is it ja jw ka kk km kn ko la lb ln lo lt lv mg mi mk ml mn mr ms mt my ne nl nn no oc pa pl ps pt ro ru sa sd si sk sl sn so sq sr su sv sw ta te tg th tk tl tr tt uk ur uz vi yi yo yue zh

Accuracy follows Whisper's own training data: excellent for high-resource languages (English, Spanish, Japanese, …), weaker for low-resource ones. That is what hybrid mode is for — a community fine-tune for one language can be plugged in with one line in babelscribe/models.py (Thai and Hindi so far). PRs adding fine-tunes for other languages are the most valuable contribution.

Limitations (honest)

• Tested end to end so far on an AMD Radeon RX 9070 XT (Windows, Vulkan). CUDA, Linux and macOS builds are produced by CI; reports from those machines are welcome.
• Hybrid mode runs two models, so it is slower (≈1 min per minute of audio with a large fine-tune on that GPU).
• Proper nouns can still be misspelled — check names before publishing subtitles.

Features

• Any input — mp4, mkv, mov, mp3, wav, m4a… (ffmpeg is bundled through imageio-ffmpeg).
• Any language — -l auto detects it; or pass -l th, -l ja, -l es…
• Picks the right GPU — prefers a discrete card over an integrated one; babelscribe devices lists them, --device N overrides.
• Long files that don't loop — runs whisper with --max-context 0, which stops the classic "same sentence repeated forever" hallucination on long recordings.
• Hybrid mode for better spelling in your language — community fine-tunes (e.g. Thai Thonburian Whisper) spell far better but often lose timestamps. --text-model takes the text from the fine-tune and the timing from large-v3-turbo, aligned character by character (works for languages without spaces), cut only at word boundaries, with the timing model filling any words the fine-tune skipped.
• --accurate — slower, fewest errors: large-v3 with beam search, or for Thai and Hindi the best community fine-tune (hybrid). Chosen per language from the FLEURS benchmark above.
• Outputs — srt, vtt, txt, json (segments with token timings).

Usage

babelscribe talk.mp4 -f srt,vtt,txt,json          # all formats
babelscribe podcast.mp3 -l en -m large-v3          # pick language and model
babelscribe talk.mp4 -l hi --accurate              # slower, fewest errors (best model per language)
babelscribe vo.wav -l th --text-model thai-thonburian   # hybrid: pick a Thai fine-tune yourself
babelscribe devices                                # GPUs whisper.cpp can see
babelscribe models                                 # models and fine-tunes
babelscribe talk.mp4 --bin /path/to/whisper-cli    # use your own whisper.cpp build

Models download on first use to ~/.babelscribe/models (BABELSCRIBE_MODELS to change). Hybrid fine-tunes are converted on your machine from their original Hugging Face repo — install pip install "babelscribe[finetune]" once; converted weights are never redistributed, so each fine-tune keeps its own licence. Thai word boundaries: pip install "babelscribe[thai]".

Add a language fine-tune

Add one entry to FINETUNES in babelscribe/models.py (Hugging Face repo, language code, whether it needs beam search) and open a PR.

Prebuilt binaries

.github/workflows/build-binaries.yml builds whisper-cli + whisper-quantize for every row of the table above and attaches them to each v* release. Point BABELSCRIBE_RELEASES at another URL to self-host.

ภาษาไทย

ถอดเสียงจากไฟล์เสียงหรือวิดีโอได้ทุกภาษา บนการ์ดจอทุกยี่ห้อ (AMD / NVIDIA / Intel ผ่าน Vulkan, Apple ผ่าน Metal) หรือ CPU ภาษาไทยแนะนำ babelscribe ไฟล์.mp4 -l th --accurate — ข้อความจาก Pathumma Whisper (NECTEC) ที่ผิดน้อยที่สุดใน FLEURS (CER 8.9% เทียบ turbo 15.9%) + เวลาจาก large-v3-turbo หรือเลือก Thonburian Whisper เอง: --text-model thai-thonburian

Credits & licence

MIT. Built on whisper.cpp (MIT) and OpenAI Whisper models (MIT). Fine-tunes belong to their authors: Pathumma Whisper by NECTEC, Thonburian Whisper by biodatlab, whisper-hindi-large-v2 by vasista22 (Speech Lab, IIT Madras).