Audio to text

Transcribe audio to text in three steps

1. Choose an audio file, a voice memo (M4A from an iPhone works) or a video, or try the sample. The page shows the size of the one-time model download and an estimate of how long your recording will take.
2. Press Transcribe. Lines appear as each part of the recording is done; you can stop at any time and keep what is finished.
3. Click any word or timestamp to hear it, fix what you need in your editor, and copy the text or download it as .txt, .srt or .vtt subtitles.

How the transcription works

The recording is decoded in your browser, mixed to mono and resampled to 16 kHz, the format Whisper expects. Whisper reads 30 seconds at a time, so longer recordings are split into pieces of up to 28 seconds, each cut at the quietest moment near the end of the window so a word is never sliced in half. Each piece goes through Whisper base.en, OpenAI’s English model (MIT licence), running in a background Web Worker through transformers.js and ONNX Runtime Web, with word-level timestamps. The words are put back on one timeline and grouped into lines at sentence ends and pauses. Nothing about the audio leaves the tab; only the model and engine files are downloaded.

Timestamps, subtitles and click-to-play

Every word keeps the time it was spoken. In the transcript, click a word to play the recording from just before it, or click a line’s time to play from the start of the line; the word being played is highlighted as you listen. The .txt download has a [m:ss] stamp at the start of each line (switch it off for plain paragraphs). The .srt and .vtt files are subtitles: short lines of up to 42 characters and 5 seconds, broken at sentences and pauses, ready for YouTube, Premiere Pro, DaVinci Resolve, CapCut or a web video player.

Speed on your device

On an Apple M2 laptop in our tests (headless Chromium, measured on a 185-second recording), an hour of speech took about 7 minutes with WebGPU, where the encoder runs on the graphics chip, and about 13 minutes with WebAssembly alone, the path browsers without WebGPU use. Phones and older computers are slower, sometimes several times slower. The page estimates the time before you start and shows the time left while it works. The 28 MB engine and the model download once and then stay in your browser’s storage, so later recordings start at once, even offline.

Accuracy and honest limits

Whisper base.en is a small model chosen so it can run on a laptop or phone, and it is English only: other languages come out as nonsense English or nothing. On clear, close-miked speech it usually gets most words right; it struggles with music under the voice, several people talking at once, strong accents, names and jargon, and very quiet recordings. Like all Whisper models it tidies speech, dropping most ums and repeated words, unless you tick “Keep ums and uhs”. On long silences or music it can occasionally write a phrase nobody said, so read the transcript against the audio. It does not label who is speaking. Very long recordings take time and memory: a computer handles an hour comfortably, a phone may not.