Transcribe audio to text in your browser: drop an MP3, WAV, M4A or a video and the words come back as a plain transcript you can copy, correct in place, and download as .txt. Whisper runs on your own machine, so the recording is never uploaded — which is the whole point for interviews, meetings, dictation and anything covered by an NDA.
How it works
- Drop an audio file or a video. The audio is decoded on your device; nothing is sent anywhere.
- Pick the spoken language and a model — Fast (≈41 MB) or Accurate (≈77 MB), downloaded once and then cached for offline use.
- Press Transcribe. Text appears live, each line plays back on click so you can check a word, and Download .txt saves the transcript.
Questions
- Can I get timestamps or subtitles instead?
- Yes. This page opens on plain text because that is what most people come here for, but the Output setting switches to timestamped subtitles at any time, and .srt and .vtt downloads are always available next to .txt. If subtitles are the job, the subtitle generator opens on them by default.
- What file types can I transcribe?
- Anything your browser can decode as audio, which in practice means MP3, WAV, M4A/AAC, OGG, FLAC and WebM, plus the audio track of MP4, MOV, MKV and WebM video. The file is decoded locally to 16 kHz mono, which is what Whisper listens to.
- How long a recording can it handle?
- The decoded audio is held in memory, so length is bounded by the tab rather than by a server quota. Files past that budget are refused before decoding with the number stated, instead of freezing the tab — and a run in progress can be stopped at any point, keeping the text produced so far.
- Does my file get uploaded?
- No. The models and all processing run inside your browser on your own device. You can verify this yourself: open your browser’s DevTools, watch the Network tab, and process a file: there are no uploads. After the first visit most tools also work fully offline.
- How accurate is it?
- The models are OpenAI’s Whisper (tiny and base tiers). On clear speech they are very usable; heavy accents, crosstalk and noisy rooms will show errors. The Accurate tier is noticeably better on difficult audio.
- Which languages work?
- Fifteen, picked in the Language setting: English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Polish, Turkish, Japanese, Korean, Chinese, Arabic and Hindi. The decoder is conditioned on the one you pick, so it transcribes that language rather than approximating it in English — which also means picking the wrong one returns confident nonsense. The weights understand 99 languages; these are the ones we measured as good enough to offer.
- How long does it take?
- On a machine with a supported GPU, a few seconds per minute of audio. On CPU-only machines it is slower but still practical, around a quarter of the recording’s length.