Transcrape

Local video/audio transcription. Pick a file, get a transcript in any language — everything runs on-device with whisper.cpp, nothing is uploaded anywhere.

Built with SvelteKit (Svelte 5) and MoreThanUI. Ships two ways: a local web server (this page), and native apps via Tauri v2 for desktop and Android — grab those from the releases page. iOS can't be installed from a download (Apple doesn't allow it); it builds from source with a free Apple ID — see docs/MOBILE.md.

What it does

• Transcribe any video or audio file ffmpeg can open, in any language whisper supports — auto-detected, or pick one up front.
• Follow along with the file playing: the current line is highlighted and the word being spoken lights up inside it. Click a word to jump to it, or a timestamp to jump to the line.
• Edit the text inline. Edits save as you type and feed the exports.
• Export .txt or .srt, or copy the whole thing.
• Translate a finished transcript to English.
• History of every transcript, kept in the browser's localStorage.

Requirements

Node	22 or newer
pnpm	any recent version (development only)
ffmpeg / ffprobe	decodes the source file to 16 kHz mono WAV
whisper-cli	from whisper.cpp — does the transcription

# macOS — whisper runs on the GPU through Metal
brew install node pnpm ffmpeg whisper-cpp

# Debian/Ubuntu — ffmpeg is packaged, whisper.cpp has to be built
sudo apt install ffmpeg
# then build https://github.com/ggml-org/whisper.cpp and put whisper-cli on PATH

# Windows
winget install OpenJS.NodeJS Gyan.FFmpeg
# then grab a whisper.cpp release binary and set WHISPER_CLI in .env

Binaries are looked up on the PATH plus the usual package-manager prefixes (/opt/homebrew/bin, /usr/local/bin, /snap/bin, ~/.local/bin). Set WHISPER_CLI, FFMPEG or FFPROBE in .env to override.

Run it

make setup   # checks tools (make doctor) + installs dependencies
make dev     # dev server on :5173

Open http://localhost:5173 and drop in a video or audio file.

Run bare make for the full target list:

Target	Does
make doctor	check the tools and models this machine has
make setup	doctor, then install dependencies
make dev	dev server on :5173
make build	release archives for every platform, into dist/
make start	run the production server out of build/
make preview	serve the production build locally
make check	type-check with svelte-check
make release	bump the version, tag, and publish a GitHub release via CI
make clean	drop build output and caches (keeps models/)

Building for other platforms

make build                        # all five targets
PLATFORMS=linux-x64 make build    # just one

dist/ gets one archive per platform plus a SHA256SUMS:

Platform	Archive
darwin-arm64, darwin-x64	transcrape-<version>-<platform>.tar.gz
linux-x64, linux-arm64	transcrape-<version>-<platform>.tar.gz
win-x64	transcrape-<version>-win-x64.zip

The server itself is portable JavaScript — SvelteKit's adapter-node bundles the whole app into build/, and there are no native dependencies. What differs per platform is the launcher (./transcrape vs transcrape.cmd), the archive format, and the INSTALL.md shipped inside with the right install commands. So the build runs on any one machine and produces every target — that's what CI does.

Unpack an archive anywhere and run it:

tar -xzf transcrape-0.1.0-darwin-arm64.tar.gz
cd transcrape-0.1.0-darwin-arm64
./transcrape        # http://127.0.0.1:5173

Node still has to be installed, and so do ffmpeg and whisper-cli — the archives don't bundle them. HOST, PORT and BODY_SIZE_LIMIT are read from the environment or from a .env next to the launcher.

Models

Pick one in Settings → Transcription. Bigger models hear better and take longer; every one is available on every device, so a phone can run the largest if you want it to. Downloads go into models/ on demand, show progress, and resume if interrupted.

Model	Size	Translates	Good for
tiny	31 MB	yes	an old phone, or a rough first pass
base	60 MB	yes	clear speech in a quiet room
small	190 MB	yes	suggested on phones — the usable/fast balance
medium	539 MB	yes	hard audio, on a laptop
large-v3-turbo	574 MB	no	suggested on desktop — the default
large-v3	1.03 GB	yes	the most accurate, ~2× turbo's time

Sizes are the real download; all are q5 quantised, roughly a third of the float original for no audible difference in a transcript.

The turbo default is transcription-only: it silently ignores whisper's -tr flag and returns the source language. So "Translate to English" is a separate step on a finished transcript, and with turbo selected it downloads medium alongside. Choosing any other model makes translation free — it reuses the model already on disk.

Switching models leaves the old one in place, so switching back is instant. The choice is remembered in models/selected.json; WHISPER_MODEL_ID pins it for a headless install.

How it works

Videos show a poster frame, grabbed in the browser — the file isn't uploaded until you press Transcribe. Language is auto-detected (or pick one under "Spoken language"). While it runs you get a progress bar with a time estimate, and you can cancel.

Word timings come from whisper's own token offsets (-ojf), so they're approximate — good enough to follow along, not frame-accurate. Editing a line drops its word timings, since they describe the wording that was replaced; that line falls back to whole-line highlighting.

Every transcript is saved to History (in this browser's localStorage — it never leaves the machine either). Playback and translation from history re-use the decoded audio the server keeps for a few hours; that audio is what Whisper was fed, so it's sound only even for a video. After it expires the transcript stays readable and editable, but the file has to be transcribed again to play or translate it.

See docs/ARCHITECTURE.md for the layout of the code.

Configuration

Everything is optional — copy .env.example to .env and uncomment what you need.

Variable	Default
WHISPER_CLI	discovered	path to the whisper.cpp CLI
FFMPEG, FFPROBE	discovered	paths to the ffmpeg binaries
WHISPER_MODEL_ID	chosen in the UI	pin the model — see Models
WHISPER_MODEL_DIR	models	where the .bin files are kept
JOB_TTL_MS	21600000 (6 h)	how long decoded audio stays on disk
MAX_JOBS	20	how many decoded jobs are kept
HOST, PORT	127.0.0.1, 5173	production server only
BODY_SIZE_LIMIT	Infinity	upload cap, production server only

Privacy

No audio, video or text leaves the machine. The only outbound requests the app makes are the two model downloads from Hugging Face, and they can be pointed elsewhere or pre-seeded by dropping the .bin files into models/ yourself.

Contributing

See CONTRIBUTING.md.

License

MIT