All projects
Media Transcription
Hebrew-first transcription that runs on your own computer.
Status: Open source
What it is
Send a video or a recording to your Telegram bot and get back a transcript with every speaker marked. Hebrew goes to a model tuned for Hebrew, other languages to a general one. The transcription runs on your own computer, so your recordings never go to a transcription service. A page in your browser shows the progress and the downloads.
What it does
- Marks who is speaking (Speaker 1, Speaker 2), and you can turn it off
- Subtitles (SRT) and plain text
- A Hebrew-tuned model for Hebrew, and a general Whisper model for other languages or auto-detect
- A browser page in Hebrew or English, with live progress, a text preview and downloads
- A private Telegram bot that sends the transcript back, files up to 2GB with your own Telegram API keys
- A job queue that carries on after a restart
Technical details
Built with
- FastAPI
- Celery
- Redis
- SQLite
- faster-whisper
- pyannote
- React
- Docker
- A FastAPI web app with a Celery worker and Redis for the job queue. Jobs and settings are kept in SQLite.
- A React (Vite) interface, built into the same Docker image.
- Transcription with faster-whisper: ivrit.ai’s Hebrew fine-tune of Whisper large-v3 for Hebrew, and OpenAI’s Whisper large-v3 for other languages.
- Speaker labels with pyannote, through ivrit.ai’s open mirror, so no Hugging Face token is needed.
- The Telegram bot talks to a local Bot API server, which raises the file limit from 20MB to 2GB.
- Runs locally with Docker Compose. The only outside traffic is the one-time model download, and Telegram if you use the bot.
- MIT licence.