Speech to Text
Dictation using your browser's built-in recognition. Free, no account, no API key.
Keep listening
Auto punctuation
Your words will appear here.
0 words
Dictate into your browser and get text back, with no API key and no account. Handy for long-form notes, or when typing is inconvenient.
Dictation here uses the Web Speech API. On Chrome and Edge that means the browser streams the audio to Google’s recognition service and returns the transcript; on other browsers the feature may be missing entirely, and the page says so rather than failing quietly. The recording is never sent to or stored by this site, but it is worth knowing that the browser sends it somewhere.
Recognition quality depends far more on the microphone and the room than on the settings. Selecting the right language before you start matters — an English model will produce confident nonsense from Korean speech — and punctuation is not inferred, so spoken text arrives as an unbroken run that needs a pass afterwards.
Recognition arrives in two passes, which explains behaviour that otherwise looks like a bug. The API emits interim hypotheses while you are still speaking and then replaces them with a final result once the phrase settles, so words appear, change and occasionally vanish as you talk. The revision is the model using the rest of a sentence to reinterpret its beginning — the same thing a person does on mishearing a word and correcting themselves a moment later. Judging accuracy from the interim text is misleading; the final transcript is the one worth reading, and it is often noticeably better than what flickered past on the way to it.
A dictation session is not open-ended. Browsers stop recognition after a stretch of silence, and continuous mode restarts the session rather than genuinely running forever, so a long dictation is a series of sessions stitched together and anything spoken in the gap between them is simply lost. Recognition also needs a network connection on the browsers that implement it, which means it stops working entirely when you go offline even though the microphone still records perfectly well. Neither limit is announced; both show up as words that should have been captured and were not.
Errors cluster in predictable places: proper nouns, product names, technical vocabulary and acronyms, none of which a general model has any reason to expect. There is no way to supply a custom vocabulary through this API, so the realistic workflow is dictation followed by a correction pass rather than dictation on its own. Punctuation is not inferred, and mixed-language speech is recognised in whichever language was selected — an English model given Korean speech produces fluent, confident nonsense rather than failing visibly, which is worse, because nothing about the output announces that it is wrong.
Microphone access follows the browser permission model rather than this page. Permission is granted per origin and remembered, the page must be served over HTTPS for the prompt to appear at all, and the recording indicator in the tab or the system tray is the browser reporting on itself — nothing a page does can suppress it. Revoking access later is done from the padlock in the address bar or the site settings, not from here. If the prompt never appears, an existing block for this origin is the usual reason, and it has to be cleared in those settings before the page can ask again.
It relies on the Web Speech API, which is well supported in Chrome and Edge, partially in Safari, and not in Firefox by default.
This page does not send it, but the browser’s speech recognition may forward audio to the vendor’s service — that is how Chrome implements it. Avoid dictating sensitive material.
Browsers end a recognition session after a stretch of silence. Start it again to continue.