Start recording and the sentences fill the screen

There are two kinds of sentence

A recognition engine rewrites a sentence several times before it commits to it — "we'll start the meeting" arrives as "we'll", then "we'll start the meat", then the finished line. Showing that process as-is makes the screen shake too much to read.

So unconfirmed text (interim) sits as one faint line at the bottom, and only confirmed text (final) accumulates into the transcript. You read stable text, and you know someone is mid-sentence from the faint line.

The live transcript is a draft

The live path uses the browser's own speech recognition, so it costs nothing to run. In exchange it cannot tell voices apart, and its punctuation and proper nouns are weak.

When you stop recording, the saved audio is processed again from the beginning. That pass separates speakers, applies the glossary and attaches accurate timestamps. Its result replaces the draft rather than joining it — otherwise the same words would be in the record twice and counted twice in the summary.

Sentences a person edited by hand are never overwritten by reprocessing.

The transcript itself

Common questions

If the internet drops, do I lose the transcript?

No. Recognition failures are isolated so they cannot stop the recording, and audio keeps being captured. When the connection returns, recognition resumes, and the pass after you stop fills in what was missed.

How accurate is it?

Practical, for clear speech in a quiet room. Names and product terms improve a great deal if you put them in the note glossary — it is the single setting that contributes most to accuracy.