🎙️ Voice Notes

Record your voice in the browser and get a live transcript. Plain HTML, CSS and JavaScript, no build step, no backend.

Features

• Record audio with MediaRecorder
• Two transcription modes:
  • ⚡ Live — Web Speech API (Urdu, English, Hindi, Arabic), instant, needs internet
  • 🔒 Whisper offline — runs Whisper in your browser via transformers.js, fully private after the one-time model download
• Whisper model picker: whisper-small (better accuracy) or whisper-tiny (faster, smaller)
• Waveform visualizer
• Edit, copy or download the transcript (.txt)
• Download the recording

Offline (Whisper) mode

1. Pick 🔒 Whisper offline under "Transcription mode".
2. Choose a model: small (default, better accuracy) or tiny (faster, less accurate).
3. Record as usual, stop, then hit "Transcribe with Whisper".

First run downloads the model once (~550 MB for small, ~100 MB for tiny) and the browser caches it — after that, transcription works with no internet at all. A progress bar shows the download. The model runs on your device's CPU, so expect roughly real-time-to-slower speed on phones; recordings over 3 minutes show a warning because they'll be slow on the small model.

Whisper reads the language from the language selector: اردو → urdu, English → english, हिन्दी → hindi, العربية → arabic.

Privacy notes — read this

• Live mode is NOT private: in Chrome, the Web Speech API sends your audio to Google's servers for recognition.
• Whisper mode IS private (after first setup): audio never leaves your device. But the first run needs internet twice — once to fetch the transformers.js library from the jsDelivr CDN, once to download the model weights from Hugging Face. After that, airplane mode works.

Run it

Mic access needs HTTPS or localhost:

python3 -m http.server 8000
# open http://localhost:8000

Use Chrome or Edge for best results.

Notes

• In Chrome, the Web Speech API sends audio to Google's servers for recognition, so Live mode is not fully offline. Use Whisper offline mode for private transcription.
• Whisper small (~550 MB) is much heavier than tiny (~100 MB) — on a slow connection, start with tiny.
• Whisper accuracy on Urdu/Hindi is generally better than the Web Speech API, but still not perfect — always proofread.

Roadmap

• Whisper in-browser (offline mode)
• Save notes in IndexedDB
• AI summary of each recording
• Roman Urdu output

License

MIT