Audio to Text — transcription in paragraphs with speaker labels

After choosing a file, tick "Label speakers" in the "Speakers" block — a "How many people are speaking" field appears. Leave it empty and the number is estimated by ear; but if you know how many people are on the recording, type it in: on recordings with similar voices (two colleagues, siblings, people of the same age) that is noticeably more accurate than guessing. In the finished text, click a "Speaker 1" label and type a name — it is applied to every turn of that person and to the downloaded files.

It happens on very long recordings and where voices are nearly indistinguishable: then a note appears above the text saying the labels could not be placed, and the text itself is kept in full. Enter the number of participants in "How many people are speaking" and run the transcription again — with a known number the voices separate more reliably. If the recording was made on separate tracks (each person on their own channel), the tool offers to recognize each track separately — the most accurate option.

Click the time at the start of the passage: its menu lists the speakers, and choosing one changes the label for that passage only. If one passage holds two people’s words — a short answer and the next question — select those words and open the same menu: only the selection moves to the other speaker. When one person was split into two voices, give the second voice the same name and the tool asks whether to merge them. The Undo button in the “Who is speaking” block reverses any of these steps.

Above the finished text there are two switches — "Timestamps" and "Speakers". Untick the one you do not need and it disappears not only from the screen but from the copy, DOCX, PDF and TXT as well: you download exactly what you see. SRT and VTT subtitles always carry time — it is what puts each line at its second of the video.

Choose server processing: the file goes to our server, you can close the tab, and when the text is ready an e-mail arrives with a link. A copy of the result waits in "My generations", so the transcript does not vanish with a closed browser. Such files can also be processed in the browser, but that takes hours and depends on the tab staying open.

Open "Names and terms" before you start and list them separated by commas: participants' surnames, company and product names, professional vocabulary. The recognition uses the list as a hint and writes them the way you specified rather than by sound. The list is kept in your browser and filled in next time.

First of all on the recording itself: a phone lying on the table in the middle of a meeting, background music and people talking at once produce more errors than any choice of model. For Russian speech the server model makes noticeably fewer errors on phone and "room" recordings than the browser one. If the recording is muffled or noisy, run it through speech restoration first — the link is next to the settings.