Voice Typing in the Browser: What It Gets Right and Wrong
How free browser speech to text actually works, where the audio goes, and the habits that take accuracy from 80% to 95%.
I type about 70 words a minute on a good day. I talk at 140. That gap is the entire reason voice typing exists, and it’s why Speech to Text is worth ten minutes of your time even if your first attempt produces nonsense.
The feature has been in Chrome since version 25, back in 2013, under the name webkitSpeechRecognition. Thirteen years later Firefox still hasn’t shipped the recognition half of the Web Speech API, which is why this tool politely refuses to work there.
The first draft is not the final draft
Dictation is terrible for polished prose and excellent for getting words out of your head. I use it for meeting notes, first drafts of emails I’m dreading, and the rambling middle section of an article where I know roughly what I want to say but not how.
What comes out is messy. Names are wrong. The engine hears “belun” as “below in”. Sentence boundaries land in odd places. All of that is fixable in two minutes of editing — which is still faster than typing the thing from scratch.
Where dictation genuinely fails: anything with code, anything with numbers you can’t afford to get wrong, and anything you’d read aloud slowly because you’re still thinking. Speech recognition needs you to commit to a sentence. If you trail off mid-thought, it guesses, and it guesses badly.
Where your audio actually goes
This is the part most “private, browser-based” transcription pages gloss over, so here it is straight: nothing is uploaded to belun.app, but Chrome and Edge do not do the recognition on your machine. They stream short audio snippets to Google and Microsoft speech services respectively and send back text. That’s how the accuracy gets to be as good as it is on a 4 MB page with no model download.
Safari on recent Apple hardware can run recognition on-device. If you’re dictating something genuinely confidential — a medical note, an unreleased contract — use Safari, or use the dictation built into macOS, iOS, or Windows instead. For a shopping list, nobody cares.
Punctuation is a vocabulary, not a feature
The raw recognition result has no punctuation at all. Google Docs voice typing solved this by teaching people spoken commands, and the same words work here when the punctuation option is on:
- “comma”, “period” or “full stop”, “question mark”, “exclamation mark”
- “colon”, “semicolon”, “dash”, “open quote”, “close quote”
- “new line” and “new paragraph”
Russian equivalents work too: запятая, точка, новый абзац. Say the word, get the character. The tradeoff is obvious the first time you try to dictate a sentence that genuinely contains the word “period” — turn the option off for that one.
Four habits that fix most errors
Set the language correctly before you start. An English-language model listening to Russian doesn’t produce an accent problem, it produces gibberish; this is the single biggest cause of “voice typing doesn’t work”.
Get within arm’s reach of the microphone. A laptop lid mic two feet away picks up the room as much as you.
Speak in complete sentences at a normal pace. Slow, deliberate speech is worse, not better — the model uses surrounding words as context, and it has less to work with when you pause between each one.
Don’t fight the pauses. The browser engine closes the session after a few seconds of silence; the tool restarts it for you, so you can stop and think without touching anything.
Open Speech to Text, set your language, and talk for sixty seconds about what you did yesterday. You’ll know within a minute whether it’s going to earn a place in your routine.