Audio to text
Transcribe English or Malay speech, on your device.
Runs entirely in your browser. Verify it in your network tab.One file at a time.
English: 41 MB or 77 MB. Malay and other languages: 77 MB or 250 MB. Each is downloaded once.
What this does
Runs OpenAI Whisper on your device through WebAssembly. English uses the English-only models; Malay and other languages use the multilingual ones. The model is downloaded once and cached by your browser for every later run. The transcript is shown ready to copy, and saved as a .txt.
Limits
- Whisper is weaker in Malay than in English. On 12 clearly read Malay sentences from Google's FLEURS set, Fast got about 4 words in 10 wrong and Accurate about 1 in 4. Use Accurate for Malay, and check the result.
- The first run of each model downloads its weights once, from 41 MB to 250 MB. After that it works offline.
- Accuracy is well below the full-size cloud models, especially with accents, several speakers, mixed languages or background noise.
- Long recordings are slow: expect a sizeable fraction of the audio length, and much more with the Accurate model.
Is this private?
Yes, and you can check rather than take our word for it. Open your browser's network tab and run the tool: your file never appears in it, because there is no server to send it to. The speech-recognition runtime is served from this same domain. The Whisper model is downloaded once from Hugging Face: a download of the model, never an upload of your audio.
