RunToolRunSpeech to Text
Convert speech to text.
How to Convert Speech to Text
Allow microphone access
Click start and grant the browser permission to use your microphone.
Speak naturally
Your words appear as text in real time, powered by the browser's recognition engine.
Copy the transcript
Edit if needed and copy the result into any document or app.
This tool is the exception: audio leaves your device
Almost everything here runs entirely in your browser. Speech recognition does not. It uses the browser's built-in speech API, and in Chrome — which is where it mostly works — that API streams your audio to Google's servers for processing and returns the text. Safari sends it to Apple.
That is worth stating plainly rather than burying. For dictating a shopping list it is unremarkable. For a confidential meeting, a medical note, a legal conversation or anything covered by a data protection obligation, it means the recording has been handled by a third party. If that matters for your material, use offline transcription software instead.
Browser support is narrow
The speech recognition API is not evenly implemented. Chrome and Chromium-based browsers support it; Firefox does not; Safari's support is partial and behaves differently. This is unusual for a web feature in 2026 and is a consequence of the API depending on a server-side service rather than on code in the browser.
The practical effect is that the tool works or does not work depending on what you opened it in, and there is no fallback that could be written to fix that.
What determines accuracy
Audio quality dominates everything else. A decent microphone close to the speaker in a quiet room produces dramatically better results than a laptop microphone across a table. Background noise, overlapping speakers and room echo are the three things that degrade recognition fastest.
Accent and dialect matter too, and not evenly — recognition systems are trained on data that over-represents some accents and under-represents others, so error rates differ between speakers of the same language. Setting the recognition language correctly, including regional variant where offered, is the cheapest accuracy improvement available.
Punctuation, names and numbers
Recognition produces a stream of words; punctuation is inferred, and inference is unreliable. Long spoken passages frequently arrive as a single run-on sentence, and speaking the punctuation aloud is the usual workaround.
Proper nouns are the other systematic weakness. Names, technical terms, product names and abbreviations are exactly the words least likely to be in the model's vocabulary, and they are also the words where an error changes the meaning rather than the phrasing. Numbers and units are similar: 'fifteen' and 'fifty' are acoustically close and semantically far apart.
Treating the output as a draft
Dictation is fastest when you accept that it produces a first draft rather than a finished text. Speaking at a steady, moderate pace in complete phrases gives the recogniser more context to work with than speaking word by word, which counterintuitively makes accuracy worse.
Always read the result before sending it. The characteristic failure is not gibberish, which you would notice, but a plausible wrong word in an otherwise correct sentence — which you will not.