Skip to main content
Glama

babelscribe — free, offline speech-to-text and subtitles on any GPU (AMD, NVIDIA, Intel, Apple)

PyPI License: MIT MCP server

babelscribe turns any audio or video file into subtitles (SRT, VTT) and text, in any of Whisper's 99 languages, on your own graphics card — AMD Radeon, NVIDIA, Intel Arc or Apple Silicon — or the CPU. No CUDA needed, no cloud, free (MIT). It runs OpenAI Whisper through whisper.cpp with Vulkan, CUDA or Metal, and works from the command line, from Python, or from AI apps (Claude, Codex, Antigravity, Gemini CLI, Cursor) as an MCP server.

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.

Your hardware

Backend used

Prebuilt asset

AMD / NVIDIA / Intel GPU on Windows

Vulkan

windows-x64-vulkan

NVIDIA on Windows

Vulkan (same build)

windows-x64-vulkan

NVIDIA on Linux (alternative)

CUDA

linux-x64-cuda (--flavor cuda)

AMD / NVIDIA / Intel GPU on Linux

Vulkan

linux-x64-vulkan

Apple Silicon

Metal

macos-arm64-metal

No usable GPU

CPU

*-cpu (--flavor cpu)

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.

Use it from an AI app (MCP) — no terminal needed

babelscribe is also an MCP server, so any AI app that supports local MCP servers can transcribe files on your computer with your GPU. Ask in plain words — "make Thai subtitles for the newest video in my Downloads" — and the app finds the file, transcribes it, writes .srt / .txt next to it, and can then proofread names, translate or summarise the text for you.

Tools: transcribe (file → subtitles + text), transcribe_folder (every audio/video in a folder, one summary back), find_media (newest audio/video in Downloads, Videos, Desktop, …), list_devices, list_languages_and_models.

transcribe_folder goes on after a file fails and lists the failure in the summary. Subtitles land next to each file, or under output_dir with the same sub-folders. Two files that would share a subtitle name (talk.mp4 and talk.mp3) keep their full name: talk.mp4.srt, talk.mp3.srt.

Claude Desktop (one click): download babelscribe.mcpb from the latest release and double-click it (or Settings → Extensions → Install extension).

Everything else runs the same command — uv fetches babelscribe for you: uvx babelscribe mcp

App

How to add it

Claude Code

claude mcp add babelscribe -- uvx babelscribe mcp

OpenAI Codex CLI

codex mcp add babelscribe -- uvx babelscribe mcp, then in ~/.codex/config.toml under [mcp_servers.babelscribe] set tool_timeout_sec = 3600 and startup_timeout_sec = 120 (defaults are 60 s / 10 s — too short for a long video or the first model download)

Google Antigravity

agent panel → ⋯ → MCP Servers → Manage MCP Servers → View raw config, add the JSON below

Gemini CLI

add the JSON below to ~/.gemini/settings.json

Cursor

add the JSON below to ~/.cursor/mcp.json

VS Code (Copilot)

.vscode/mcp.json: same entry under "servers" with "type": "stdio"

{
  "mcpServers": {
    "babelscribe": { "command": "uvx", "args": ["babelscribe", "mcp"], "timeout": 3600000 }
  }
}

(timeout is in milliseconds and only some apps read it.) Already installed with pip? Use "command": "babelscribe", "args": ["mcp"]. Web-only chat apps (e.g. grok.com, chatgpt.com) can only reach servers on the internet, not your computer, so they can't use your GPU or files.

Related MCP server: video-transcriber-mcp

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).

Language

Talk

Length

default: time / error

--accurate: time / error

English

Matt Walker — Sleep Is Your Superpower (TED)

19.3 min

48 s (24x) · WER 1.5%

107 s (11x) · WER 1.5%

Japanese

Kazunari Taguchi (TEDxHimi)

16.2 min

47 s (20x) · CER 4.7%

115 s (8x) · CER 4.2%

Spanish

Adrià Solà Pastor — Cómo hablar (TEDxESIC University)

19.9 min

56 s (21x) · WER 13.8%

134 s (9x) · WER 13.5%

Thai

นิติ ชัยชิตาทร — โปรดเรียกฉันด้วยนามอันแท้จริง (TEDxBangkok)

14.1 min

74 s (12x) · CER 22.7%

565 s (2x) · CER 16.4%

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.

Language

default (turbo)

--accurate

model --accurate uses

--accurate, edge-trimmed*

English

6.1%

5.7%

large-v3, beam 5

5.6%

Spanish

4.1%

4.2%

large-v3, beam 5

2.6%

French

7.4%

7.3%

large-v3, beam 5

7.2%

German

4.3%

4.0%

large-v3, beam 5

3.3%

Portuguese

8.5%

7.7%

large-v3, beam 5

5.3%

Italian

6.7%

6.5%

large-v3, beam 5

3.8%

Russian

6.8%

5.8%

large-v3, beam 5

5.6%

Arabic

11.0%

10.6%

large-v3, beam 5

9.7%

Hindi

28.4%

12.6%

vasista22/whisper-hindi-large-v2 + turbo timing

11.1%

Indonesian

9.8%

7.8%

large-v3, beam 5

5.6%

Vietnamese

10.8%

9.1%

large-v3, beam 5

9.1%

Turkish

6.2%

6.7%

large-v3, beam 5

5.8%

Japanese (CER)

6.4%

5.7%

large-v3, beam 5

4.6%

Chinese (CER)

6.3%

5.3%

large-v3, beam 5

4.8%

Korean (CER)

4.1%

3.9%

large-v3, beam 5

3.1%

Thai (CER)

15.9%

8.9%

Pathumma Whisper (NECTEC) + turbo timing

8.8%

* 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)

GPU on AMD / Intel

Windows, no build step

Video in, subtitles out

Long files don't loop

Better text for your language

babelscribe

✅ Vulkan

✅ prebuilt whisper-cli downloaded for you

✅ ffmpeg bundled

✅ --max-context 0 by default

✅ hybrid fine-tune text + turbo timing

whisper.cpp (raw)

✅ Vulkan — if you compile it

❌ official releases ship no Windows Vulkan build

❌ WAV 16 kHz only

⚠️ you must know the flag

❌

faster-whisper / WhisperX

❌ GPU = NVIDIA CUDA only

✅ pip

✅

⚠️

⚠️ manual

Cloud APIs

n/a (cloud)

✅

✅

✅

❌ — and your audio leaves your machine, paid per minute

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.

FAQ

How do I run Whisper on an AMD GPU on Windows? pip install babelscribe, then babelscribe video.mp4. It downloads a Vulkan build of whisper.cpp that runs on AMD Radeon (and Intel Arc / NVIDIA) cards on Windows and Linux — no ROCm, no CUDA, no compiling.

How do I make subtitles (SRT) from a video for free, offline? babelscribe video.mp4 -f srt writes video.srt next to the video. Nothing is uploaded; it runs on your own computer.

What is the most accurate free transcription for Thai? babelscribe video.mp4 -l th --accurate — Pathumma Whisper (NECTEC) for the text plus Whisper turbo for timing: CER 8.9% on Google FLEURS vs 15.9% for plain Whisper turbo. Thonburian Whisper is available too (--text-model thai-thonburian).

Can Claude / ChatGPT Codex / Gemini transcribe a video on my computer? Yes — add babelscribe as an MCP server (see Use it from an AI app). Claude Desktop installs it with one click from babelscribe.mcpb. The AI app can then find a file, transcribe it on your GPU, and proofread, translate or summarise the text.

How fast is it? About 20x real time with Whisper large-v3-turbo on an AMD Radeon RX 9070 XT: a 19-minute talk in 48 seconds.

Which languages are supported? All 99 Whisper languages, with automatic language detection. FLEURS error rates for 16 of them are in the benchmark table.

Is it better than faster-whisper or WhisperX? Those are excellent on NVIDIA GPUs; on AMD / Intel GPUs they fall back to the CPU. babelscribe's niche is any GPU, zero setup, and better Thai / Hindi through community fine-tunes. If you have an NVIDIA card and like Python, faster-whisper is a fine choice.

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.

  • Custom vocabulary — --prompt "Pitakpong, babelscribe" gives whisper an initial prompt that biases the spelling of names and jargon. In the MCP transcribe tool the same thing is the vocabulary argument. While a prompt is set, --max-context 0 is not passed (whisper.cpp ignores the prompt otherwise).

  • Translate to English — --translate writes the transcript translated to English (whisper's translate task, any source language). It uses large-v3 because turbo was not trained to translate, and cannot be combined with --text-model. In the MCP transcribe tool it is the translate argument.

  • Outputs — srt, vtt, txt, json (segments with token timings), ass (karaoke: each word highlights as it is spoken, from the token timings — for short-form video and players that read ASS).

Usage

babelscribe talk.mp4 -f srt,vtt,txt,json          # all formats
babelscribe reel.mp4 -f ass                        # karaoke subtitles, one highlight per word
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 talk.mp4 --prompt "Pitakpong, babelscribe"   # bias the spelling of names and terms
babelscribe talk.mp4 --translate                   # subtitles translated to English
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

From Python:

from babelscribe.api import transcribe_file
r = transcribe_file("talk.mp4", lang="auto", formats=["srt", "txt"])   # or accurate=True, prompt="names, terms"
print(r["lang"], r["device"], r["files"]); print(r["text"][:200])

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]".

Docker

No Python, ffmpeg or whisper.cpp on the host — the image has them, and the model is cached in a volume. There is no published image yet, so build it from a checkout, then run it on the folder that holds your media:

docker build -t babelscribe -f docker/Dockerfile .
docker run --rm -v "$PWD:/media" -v babelscribe-models:/models babelscribe talk.mp4   # writes talk.srt next to the input

CPU by default, linux/amd64; a Vulkan image for GPU hosts is one --build-arg away. Details: docker/README.md.

Roadmap

What's planned next (subtitle line wrapping, MCP progress for long files, folder batches, translate-to-English, more fine-tunes) is in ROADMAP.md and the open issues.

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

ใช้ผ่าน Claude Desktop ได้โดยไม่ต้องพิมพ์คำสั่ง: โหลด babelscribe.mcpb จากหน้า Releases แล้วดับเบิลคลิก จากนั้นพิมพ์ในแชทว่า "ทำซับไทยให้คลิปล่าสุดในโฟลเดอร์ดาวน์โหลด" 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

Privacy Policy

babelscribe runs entirely on your own computer. Last updated 2026-10-06.

  • Data collection: none. babelscribe has no telemetry, analytics, accounts or crash reporting, and never uploads your audio, video, transcripts or file names anywhere.

  • What it processes and where: the media file you choose is converted and transcribed locally; subtitles and text are written to your disk (next to the file, or the folder you choose). Through MCP, the transcript text is returned to the AI app that called the tool — what that app does with it is governed by that app's own privacy policy.

  • Network access (downloads only): on first use it downloads the whisper-cli program from this project's GitHub releases, speech models from Hugging Face (huggingface.co/ggerganov/whisper.cpp, and for --accurate Thai/Hindi the fine-tune's own repository), and Python packages from PyPI when installed with pip/uv. These are plain downloads; no personal data is sent. GitHub, Hugging Face and PyPI see a normal download request (IP address, user agent) under their own privacy policies.

  • Storage and retention: downloaded programs and models are cached in ~/.babelscribe (or BABELSCRIBE_HOME / BABELSCRIBE_MODELS) until you delete that folder. Outputs stay wherever they were written until you delete them. babelscribe keeps no other data.

  • Third-party sharing: none.

  • Contact: open an issue at https://github.com/phonology024/babelscribe/issues

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).

Available Tools

4 tools
find_mediaFind audio/video filesA
Read-only

Find audio/video files on this computer, newest first. Searches Downloads, Videos, Desktop, Music and Documents (two levels deep) unless folder is given. Use it when the user names a file without a full path.

Inputs: limit, folder, name_contains. Returns: result

list_devicesList GPUsB
Read-only

GPUs whisper.cpp can use on this computer and which one babelscribe picks automatically.

Inputs: no named parameters.

list_languages_and_modelsList languages and modelsC
Read-only

Models available, and which model --accurate uses per language.

Inputs: no named parameters.

transcribeTranscribe audio/video to subtitlesA
Idempotent

Transcribe a local audio or video file into subtitles / text.

file_path: absolute path to the media file (mp4, mkv, mov, mp3, wav, m4a, ...). language: ISO code such as en, th, ja, es, hi — or "auto" to detect. accurate: slower, fewest errors (large-v3 with beam search, or the best fine-tune for Thai / Hindi). formats: comma list from srt, vtt, txt, json, ass (ass = karaoke word highlight). output_dir: folder for the output files; empty = next to the media file. vocabulary: names and terms to spell correctly, e.g. "Pitakpong, babelscribe" (whisper initial prompt); empty = none. translate: true = write the transcript translated to English (uses large-v3; slower than turbo). Returns the files written, detected language, device used, and the transcript text.

Inputs: file_path (required), formats, accurate, language, +3 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.7/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a clearly distinct role: transcribe performs the core action, find_media resolves paths, list_devices reports GPU availability, and list_languages_and_models reports model/language info. The two list_* tools are separated by resource (devices vs. models/languages), so there's no realistic confusion.

    Naming Consistency4/5

    Three tools follow a snake_case verb_noun pattern (find_media, list_devices, list_languages_and_models) while the core action is a bare verb (transcribe). This is a minor, readable deviation rather than a convention clash.

    Tool Count5/5

    Four tools for a focused transcription server is well-scoped: one action tool plus three small discovery/introspection helpers. Nothing feels redundant, and each tool earns its place.

    Completeness4/5

    The core workflow is well covered — file discovery, device selection, model/language lookup, and transcription with formats, translation, and vocabulary hints. Gaps remain for job progress/status on long files and batch transcription of multiple media files, which agents can only work around by calling transcribe repeatedly.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers