Audio to Text (Speech Recognition)

How does browser-based transcription work?

Speech recognition used to mean uploading your audio to a cloud service. This does not. It runs Whisper, a well-regarded open speech-recognition model, compiled to run through WebAssembly with Transformers.js. Your browser decodes the audio, resamples it to the 16 kHz mono the model expects, and the model produces text with timestamps — all on your device. The only thing fetched from the network is the model’s weights, downloaded once from a public CDN and then cached for next time.

Is my audio uploaded anywhere?

No. This is the whole point of doing it in the browser. Your recording — which might be a private interview, a therapy session, a confidential meeting or unreleased content — is transcribed locally and never sent to a server. You can verify it: open your browser’s DevTools, watch the Network tab, and transcribe a file. You will see the model download (the first time only), and no upload of your audio.

Which model should you choose?

Whisper comes in sizes that trade speed for accuracy. Because everything runs on your CPU in the browser, smaller is faster:

Can it make subtitles, not just text?

Yes. Whisper returns timestamps for each chunk of speech, so the tool can build a real subtitle file, not just a wall of text. Export as SRT or VTT and you get timed cues ready to drop into a video player, YouTube, or an editor. From there you can fine-tune them in the subtitle editor, fix any offset with the subtitle shifter, or burn them into the footage with add subtitles to a video. For a clean reading transcript instead, export plain text or run the result through subtitle to text.

How accurate is it?

Whisper is strong on clear speech and handles accents and background noise better than older engines, but no automatic transcription is perfect. Expect the odd wrong word, especially with names, jargon, crosstalk or poor audio. Treat the output as a fast first draft you tidy up — which is still far quicker than transcribing from scratch. Clearer recordings and the Base model give the best results.