Voice note to text, free

How does a recording become a note?

In three steps. First the recording is converted to 16 kHz mono audio - the format the speech-recognition model works on, and one that a dictaphone m4a, an opus saved from a messenger, an mp3, a wav and the audio of a video all fit into. Then the audio is run through the open Whisper model in thirty-second windows, which returns not just the words but when they were said. Finally the fragments are joined into paragraphs rather than subtitle lines: we break where there is at least a second and a half of silence, or where a sentence ends and the paragraph is already long enough.

That paragraph grouping is the difference between a subtitle and a note. A subtitle line is six words long because that is what fits on screen; thirty such lines in a note are unreadable. That is also why there is no timestamp anywhere in the export: you paste the TXT or the Markdown file into your notes app and it looks as if you had typed it. The timings stay in the editor, where they have a job - click a paragraph and playback starts there.

All of this happens on your own machine. The recording never reaches a server, and neither does the text: once the page has loaded, the processing would finish even offline. The price is the first run - the model has to be downloaded, and it is several tens of megabytes. After that it stays in your browser cache and the next recording does not pay it again.

What it is good for

Dictated notes. You talk through a train of thought while driving or walking, and here it becomes searchable, copyable text. The recording stays on your phone; the text is made in your browser.

Meetings and interviews. An hour of recording becomes a rough transcript in a few minutes, one you can actually read through. We do not separate speakers - if you need that, you mark the paragraphs yourself, which is exactly why the editor lets you rewrite every one of them.

Voice messages. A long voice message is faster to read than to listen to - and it works where you cannot play audio at all. Opus and m4a files saved from messengers can be dropped straight in.

Where its limits are

Honestly: this is the smallest Whisper model, because it is the only one that fits in a browser. It is very good on clear, close-up English speech and noticeably weaker on Hungarian - it often gets proper nouns and jargon wrong, and punctuation is hit and miss. The transcript is raw material, not finished text; that is why every paragraph is editable.

A noisy room, people talking over each other, a distant microphone and a strong accent all make the result worse. We do not separate speakers, and we do not summarise. Speed depends on your machine: five minutes of audio is twenty to thirty seconds on a current desktop, several times that on a phone.

Frequently asked questions