Speech to text

How it works

Drop an audio or video file onto the box, or press Record and speak. The sound is converted on your device into a 16 kHz mono track, cut into pieces at natural pauses, and transcribed piece by piece. Each sentence appears as soon as its piece is done, so on a long recording you can start reading while the rest is still being worked on.

The first time you transcribe, the browser downloads the speech engine once. The progress card shows the real size and speed. After that it is stored by the browser and the next file starts right away.

Your recordings stay on your computer

There is no upload step. The file is read and transcribed inside the browser tab, and the text goes straight into the editor. Interviews, patient notes and client calls never leave the device.

Recordings made with the Record button are kept in memory only and are gone when you close the tab. The transcript text is saved in your browser so a reload does not lose it.

Fix the text while you listen

Every line starts with a small time chip. Click it and the original audio plays from that sentence, so checking a name or a number takes one click. While the audio plays, the line being spoken is marked in the editor.

Edit the text like any document. The times stay attached to their lines, so the SRT and VTT files you download afterwards carry your corrections with the original timing.

Fast or Accurate

Fast is the default. For English it uses an engine built only for English that works on the actual length of each piece instead of padding every piece to 30 seconds, so short recordings come back quickly. Other languages use a compact multilingual engine.

Accurate uses a larger multilingual engine and a bigger one-time download. In our tests on a laptop, with the processor doing all the work, a 22-second English recording took about 1 second in Fast and 13 seconds in Accurate, and both got the names, dates and the dollar amount right. When a name or an accent keeps coming out wrong in Fast, try Accurate before fixing it by hand.

Silence is not turned into words

Speech engines tend to invent text where nobody is speaking. In our tests a 32-second clip of wind and traffic came back as "Thank you." Long silences are therefore never sent to the engine, and short lines such as "Thank you" or "Thanks for watching" are removed only when the audio under them does not look like speech. A real "thank you" at the end of a talk stays.

What you can download

TXT gives you plain paragraphs: sentences are joined, and a new paragraph starts at pauses of 1.5 seconds or longer. SRT and VTT are subtitle files with at most two lines of 42 characters per caption, ready for YouTube, Premiere, DaVinci Resolve or CapCut. Word (DOCX) opens in Word, Pages and Google Docs.

To hear the text read back, use Read aloud in the download menu. It opens TextToSpeech.tools with the transcript already filled in.

Limits

One file can be up to 3 hours long. At 16 kHz mono an hour of audio takes about 115 MB of memory, so very long files on a phone may run out of memory before they finish. If that happens, split the file and transcribe it in parts.

Video files work when the browser can decode their sound track. MP4 with AAC audio, MOV, WebM and MKV with Opus are the most common. A file with no sound track gets a clear message instead of an empty transcript.