Speech to Text

Speech to Text turns your microphone into a keyboard. Click once, speak naturally, and finished phrases appear in an editable transcript as the browser recognises them, with not-yet-final words shown dimmed above. It is useful for hands-free note taking, drafting, accessibility, and checking pronunciation in a foreign language — and because it runs on the Web Speech API, there is nothing to install.

How to Use

1. Choose the recognition language. Pick from the 12-language dropdown. The default follows the page language, but you can change it at any time.
2. Start dictation. Click “Start dictation” and grant microphone access when the browser asks.
3. Speak and watch. Interim words appear dimmed in the live line; once the engine commits a phrase it is appended to the transcript below.
4. Stop when finished. Click “Stop dictation.” The transcript remains on the page.
5. Edit, copy, or download. Fix mistakes in the text area, then use Copy or Download .txt. Word and character counts update as you type.

What the Live View Shows

Two text regions update simultaneously. The interim line above the transcript holds the phrase the engine is still guessing at; it changes as the engine refines its hypothesis, which is why it can flicker mid-sentence. The transcript area only receives committed phrases, so it grows in stable chunks rather than rewriting itself. Word and character counts in the tool’s own counter use a rule suited to dictation: Latin words are counted by whitespace, while each CJK ideograph or kana character counts as its own word, because those scripts are written without spaces.

Because interim text is never committed, a word you see in the dimmed line is not yet part of the transcript. If you stop dictation while a phrase is still interim, that phrase may be lost — finish the thought before clicking Stop, or let the engine commit it first. The engine commits more readily at natural pauses, so a brief breath at the end of a sentence both improves accuracy and locks the text in.

Transcript Merging and Spacing

When a final phrase arrives, it is not simply concatenated. The merger inserts a single space between the existing transcript and the new segment, with two exceptions. If the new segment begins with closing punctuation — a comma, period, question mark, colon, bracket, or the CJK equivalents — no space is added, so you get “hello, world” rather than “hello , world.” If the existing text ends in a CJK character, no trailing space is added either, keeping Chinese and Japanese output continuous. Empty and whitespace-only segments are skipped entirely, so a silent recognition event never leaves a stray gap.

This matters because recognition engines emit segments at unpredictable boundaries. The same spoken sentence might arrive as one chunk or as three, and a naive textarea append would produce double spaces or a space before every comma. The merge rules normalise that away, which is why the transcript usually reads cleanly enough to copy without a formatting pass. Editing by hand still overrides everything: typing into the textarea is treated as ordinary text, and the next final segment is appended to whatever is there, including your corrections.

Supported Languages and Browsers

Recognition is handled by the browser, which is why support is uneven. Chrome and Edge offer it on desktop and Android; Safari supports it from 14.1 onward and may process some audio on-device; Firefox has no SpeechRecognition implementation at all, so the tool replaces the interface with a fallback notice. The 12 curated languages cover the most common use cases, but accuracy varies by language and by how closely your accent matches the training data. Proper nouns, acronyms, and names are the usual casualties.

Two practical notes on language selection. First, the selector is the recognition language, which controls the acoustic and language model used — it does not translate, so choosing German while speaking English produces German-model guesses at English words, not a German translation. Second, switching the selector mid-session affects the next recognition run; if you change it while listening, stop and start again so the new language takes effect. The same audio can produce very different transcripts under different language models, so set it before you begin.

Accuracy, Privacy, and Limits

Speech recognition returns the most probable words, not the correct words. Homophones (“their” and “there”), technical vocabulary, and heavy background noise all cause errors that you should expect to fix by hand. Punctuation is only sometimes inserted, and never reliably. On privacy: this site does not upload or store anything, but that does not mean recognition is local. In Chrome and Edge the audio is sent to the browser vendor’s service for processing, so treat dictation of confidential material the way you would treat any cloud transcription. Check your browser’s own privacy documentation if that matters for your use case.

Practical Tips

• Speak in complete phrases and pause briefly at sentence boundaries; the engine commits more accurately on natural chunks than on one unbroken stream.
• Dictate punctuation as words if you need it (“comma,” “period,” “new line”) and expect to clean it up.
• Use a headset microphone in noisy rooms — the recogniser, not the room, is choosing the words.
• After a long session, skim for recurring misrecognitions (a name spelled three ways) and use find-and-replace in the text area before exporting.
• If the session stops unexpectedly, press Start again; the existing transcript is not lost.

Related Tools

• Text to Speech — the reverse direction: turn written text into spoken audio.
• Voice Recorder — record microphone audio and download it as a file.
• Word Counter — detailed statistics for your dictated text.