Audio to text converter

The page hands your file to the browser’s own audio decoder, the same one that plays sound on any website. Whatever it can decode, the speech model can read.

The model listens to a 16 kHz mono version of your file, close to telephone quality. That is why bitrate and sample rate barely matter, and why the recording itself matters a great deal.

• Distance. A phone on the table half a metre from the speaker beats an expensive microphone across the room.
• One voice at a time. Overlapping speech is where every speech model loses words.
• Background. Music, traffic and fans turn into wrong words, or into text that was never said. Stretches of pure silence are skipped on purpose, because Whisper tends to invent sentences for them.
• Tell it the language. Automatic detection listens to the first thirty seconds. If a recording starts with music or with another language, choose the language yourself.

Names, brands and technical terms are the usual leftovers. Fix them once in the editor before exporting, or tidy dictated text on the dictation page.

Copy text gives you plain paragraphs for an email or a document. TXT and DOCX keep the start time of each line, which is what you want when quoting an interview or citing a lecture. SRT and VTT are subtitle files; the SRT generator page explains how to time them well.