How to transcribe audio to text
- Pick the Language that's spoken and a model: Fast for a quick draft, Accurate for fewer mistakes.
- Drop in an audio or video file, or press Record from microphone. The first time, the speech model downloads; then the text appears piece by piece with timestamps.
- Click a timestamp to hear that part, then copy the text or download it as TXT, or as SRT or VTT subtitles.
Which model to pick
Both settings run Whisper, OpenAI’s open-source speech recognition model, in 8-bit versions small enough for a browser. Fast is Whisper tiny, Accurate is Whisper base, which makes noticeably fewer mistakes at about half the speed. With English selected, the English-only versions are used, which are more accurate for English than the multilingual ones of the same size. Choose another language, or Detect language, for everything else.
Getting a better transcript
- Clear audio matters most. A mic close to the speaker, little echo and no music under the voice give far better results than any setting.
- Set the language rather than relying on detection, especially for short clips.
- Split long files. Transcription runs piece by piece, so a long file works, but on a phone it’s safer to trim it into parts of 20 to 30 minutes.
- Check the text. Whisper sometimes repeats a phrase or writes a plausible word it didn’t hear, mostly over music or long silences. Click the timestamp to listen and correct it.
TXT, SRT and VTT
TXT is plain text, with or without timestamps (untick Timestamps to get running paragraphs). SRT is the most widely supported subtitle format, accepted by YouTube, Vimeo, Premiere Pro, DaVinci Resolve and VLC. VTT (WebVTT) is the format for HTML5 video on websites. To go the other way, from text to speech, use the AI voice generator.