otranscribe 0.1.0

Transcribe any audio or video via OpenAI STT. Supports diarisation and cleaned rendering.

otranscribe

otranscribe is a tiny command line interface for turning any audio or video into text. It primarily wraps the OpenAI speech-to-text API, but also includes two offline backends so you can avoid network calls and API costs entirely. The CLI handles all of the boilerplate: it extracts audio from arbitrary input, normalises it, runs the transcription on your chosen engine and optionally renders a cleaned transcript with timestamps and speaker labels.

Quick start

pip install otranscribe
export OPENAI_API_KEY="sk-..."
otranscribe -i audio.mp3

See INSTALLATION.md for detailed setup including offline engines.

Features

• Any input format – as long as ffmpeg can read it, it can be transcribed.
• Diarisation support – by default it uses the gpt-4o-transcribe-diarize model and requests diarized_json output so that speakers are labelled. When you don't need diarisation or want to avoid API costs, you can select the local Whisper engine.
• Clean rendering – remove filler words, collapse whitespace and insert timestamps every N seconds and on speaker change.
• Raw output – choose --render raw to write the exact response from the engine (JSON, text, SRT, VTT, etc.).
• Choice of engine – use the OpenAI API (--engine openai) for high-quality diarised transcripts or choose one of the offline backends when you want to work without an internet connection:
  • Local Whisper (--engine local) – runs the reference openai-whisper model on your machine. This backend produces accurate transcriptions but can be relatively slow on CPU and does not assign speaker labels.
  • faster-whisper (--engine faster) – uses the faster-whisper reimplementation based on CTranslate2. It is up to four times faster than the original open source Whisper implementation and uses less memory, with optional quantisation and GPU acceleration for even greater speed. Since diarisation is not available locally, the engine assigns all words to a single speaker (Speaker 0).
• Minimal dependencies – uses requests instead of the heavy openai client when talking to the API. The local engine only imports Whisper if you choose --engine local. The faster engine pulls in the faster-whisper package only when selected.

Documentation

• INSTALLATION.md – Install, uninstall, and development setup (including pre-commit hooks and code style)
• USAGE.md – How to use otranscribe, examples, and common workflows
• TROUBLESHOOTING.md – Common errors, diagnostics, and solutions

Contributing

Contributions are welcome. Please open issues or pull requests on the project repository. See INSTALLATION.md for development environment setup.

License

Released under the terms of the MIT license. See LICENSE for details.