FreeScribe

Privacy-first, browser-based speech transcription and translation.

FreeScribe records or uploads audio, runs Whisper locally in a Web Worker, and optionally translates the resulting transcript locally with NLLB. It is currently an alpha prototype and is not intended for clinical, legal, or other safety-critical use.

Run locally

npm install
npm run dev

Useful checks:

npm test
npm run lint
npm run build
npm run evaluate:fixture
npm run evaluate:translation
npm run summarize:performance

Evaluation commands also accept --output evaluation/results/report.json to save JSON reports locally; generated results are ignored by Git.

The first transcription or translation downloads model files and may require significant memory and bandwidth.

Troubleshooting first use

• Keep the tab open while the model downloads; later runs reuse the browser cache.
• Use Xenova/whisper-base for better multilingual accuracy and Xenova/whisper-tiny when speed matters more.
• If automatic language detection is uncertain, select the spoken language manually before retrying.
• For a clean reset, open Projects and choose Clear downloaded model files, then reload the page.

Documentation

• Architecture
• Privacy
• Model card
• Product roadmap
• Security policy
• Contributing
• Performance
• Code of conduct
• Governance
• Getting help
• License

Current limitations

• Multilingual transcription defaults to the more accurate Xenova/whisper-base with local automatic spoken-language detection. Xenova/whisper-tiny remains available as a faster, lower-accuracy option. The UI exposes major Indian/South Asian languages (including Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Urdu, Kannada, Malayalam, Punjabi, Odia, Assamese, Nepali and Sanskrit) plus major world languages. English-only models intentionally restrict language selection.
• Automatic detection loads a separate local speech-language model and analyzes up to the first 15 seconds; first-use download time can be substantial, while warm-cache detection is much faster. The result screen reports the detected language and confidence. If the strongest detector result is unsupported, FreeScribe asks for manual selection instead of silently choosing a weaker guess.
• Translation quality and supported source/target language pairs depend on the selected NLLB model configuration.
• Browser support, model performance, and accuracy vary by device and audio quality.
• No server-side storage is provided.
• Saved projects use browser IndexedDB and remain on the current device/browser.
• Audio input is limited to 200 MB and 30 minutes per recording.

Never commit real recordings, patient information, or private transcripts.

License

FreeScribe's original source code is licensed under the MIT License. The Whisper and NLLB model files downloaded at runtime are separate dependencies and remain subject to their respective model licenses and usage restrictions.