Dictation that stays on your device

How the live draft follows your voice

Dictation here is not word-by-word typing. While the tape runs, the page cuts what you have said into the same 30-second windows an ordinary transcription would use. It transcribes each window as soon as it is complete.

That is why the first text appears about 15 seconds after you start, and why afterwards the draft trails your voice by about 30 seconds. You speak, a pause later the words arrive.

The draft is built from the same windows and the same merge as a transcript made after the fact.

The model downloads when you press Record, so on a first visit the opening text appears only once that download is done. Your browser keeps the model afterwards.

The recording does not depend on the draft

The recording and the draft are separate. The draft only listens to the tape and cannot change it. If the draft stops, the tape keeps running, and a failure in the speech engine is reported in the draft's own panel.

On a slower device the draft falls behind. The panel tells you it is running behind, and the rest of the text arrives after you press Stop. Nothing you said is lost because the draft could not keep pace.

While you speak, the page shows that the tape is running. You see the word REC, a lamp, a counter and a level bar that moves with what the microphone hears. A muted input shows up in the first seconds, not after the whole session.

The browser asks for the microphone when you press Record and not earlier.

When you press Stop

The draft settles into an ordinary transcript. That transcript is ready to be checked: press Play and the text follows the recording, press a line's time to jump to it, and fix words in place.

This step matters. Dictation that cannot be checked is only a faster way to misquote yourself, so the page makes checking against the recording the next thing you do.

Then export what is on screen as plain text, Markdown, a Word file, a numbered PDF, or subtitles. Corrections and any speaker names you typed are included.

Nothing is kept unless you ask. Keep the recording on this device, save the transcript to an optional account, or download it. Reload without doing any of these and the draft is gone.

When this kind of dictation fits

Notes to yourself. Think aloud while the draft builds, then tidy the text afterwards. A thirty-second lag does not matter when you are the only reader.

A first draft of prose. Say a paragraph, press Stop, then correct it. The draft is a starting point to edit, not finished writing.

A sensitive subject. The speech model runs in your browser, so your audio is not sent to a recognition service.

Not a live caption tool. A delay of about half a minute is too long for subtitling a speaker as they talk. For that job you need a different product.

Before you press Record

Decide whether you want the live draft. It is ticked on this page, under the recorder, with the label Live draft while you speak. Tick or untick it before you press Record and not after.

Press Record, speak, and watch the draft panel. Its note says when the draft is running behind. If it does, carry on talking: the tape is safe and the rest of the text arrives after you press Stop.

When you stop, read the settled transcript against the recording at least once. A draft produced live is still a machine's draft, and the recording is the record.

More questions about dictation here

Why is the draft behind my voice?

The model reads whole 30-second windows. A window can only be transcribed once it has been recorded, so the text trails by about half a minute, and further on a slower device.

Can I switch the live draft off?

Yes. The box under the recorder is ticked on this page and can be unticked before you press Record. Unticked, the recording is transcribed afterwards instead, with the same model.

Does a phone work for dictation?

It can, but a phone is slower than a computer, so the draft falls further behind. The recording itself is unaffected, and the rest of the text arrives after you press Stop.

What does the draft cost in download size?

Ticking the live draft makes Record also fetch the speech model, which is why it is a choice you make and not a default on every route.