How to transcribe an interview with speaker names

An interview transcript is only useful if you know who said what. A quote attributed to the wrong person is worse than no quote at all — it's the kind of error that ends up in print.

Automatic transcription gets you most of the way, but not by itself. What comes back is "Speaker 1" and "Speaker 2", and the line between them is a guess. This is a workflow for turning that into a transcript with real names that you can actually stand behind.

First, what "speaker labels" actually are

The feature is called speaker separation, or diarization. It doesn't recognise anyone. It groups stretches of speech by how similar the voices sound and gives each group a number. So "Speaker 1" means the first voice the system heard, nothing more — which is why you'll never get names out of it automatically, and why two similar-sounding people can end up merged into one.

Knowing that tells you where the errors will be: at the moments voices overlap, at short interjections ("right", "mm-hm") that are too brief to sound like anyone, and between speakers who sound alike. That's where you'll spend your checking time.

Before you record

Most transcript problems are decided when you press record.

1. Give each voice its own space. One microphone per person is best. Failing that, sit so each person is roughly the same distance from the one device. A voice that's much louder than the other pulls the grouping around.
2. Say names early. "Thanks for joining me, Maya" in the first minute costs nothing, and it tells you — later, listening back — which number is which.
3. Don't talk over each other more than you have to. Crosstalk is where the words get lost as well as the labels.
4. Know your language. Speaker separation isn't available in every language, and where it isn't, the transcript comes back with no speaker labels at all — there are no speakers to rename, and you'd be marking who said what by hand. Check before you record an interview whose value depends on attribution. In txscribe, the language picker tells you before you upload.

Transcribing

Choose the interview's language rather than leaving it on automatic. Automatic detection has to decide what language the speech is in, and when a recording switches — a bilingual guest, a quoted phrase — what was said in the other language may come back translated rather than transcribed, with nothing marking it.

If the recording is long, expect more drift in the labels than on a short one. Long audio is processed in pieces, and keeping the same person under the same number across pieces is the hard part; here's why a three-hour interview comes back with eleven speakers.

Turning numbers into names

This is where most of the work is, and where the tool should do the repetitive part for you.

Rename each speaker once. Change "Speaker 1" to "Maya" and every paragraph that voice spoke changes with it. In txscribe, a name you've used before is offered back the next time, so a recurring co-host is one click, not retyping.

Fix the paragraphs that went to the wrong person. Play through the transcript and listen at the boundaries — every time the label changes, and every short interjection. When a paragraph belongs to someone else, reassign it. Because every word is timed, you can click any line and hear exactly that second instead of scrubbing through the audio to find it.

Split where a turn changed mid-paragraph. Sometimes the recogniser runs two people's words into one paragraph. Split it where the second person starts, then give the new paragraph the right speaker.

Teach it the names. Proper names are what recognisers get wrong most often: an unusual surname, a company, a place. Adding them to a custom word list makes the next interview more likely to spell them right the first time. It nudges the recogniser rather than overriding it, so it cuts down the fixing without removing the need to check.

Before you quote anyone

Treat the transcript as a draft of the record, not the record. For anything that will be published or relied on:

• Listen to every line you quote, at its own timestamp, before you use it. It takes seconds.
• Check the attribution of every quote, not just the words. Label errors don't look like errors.
• Watch for lines that look complete but aren't — a garbled stretch smoothed into plausible words, a sentence cut off at a crosstalk.

That last one is the reason we built txscribe the way we did. A transcript that looks polished invites you to trust it; one where any line plays back the second it came from lets you check it.

Exporting

Once the names are right, they carry through to the export: a Word document or plain text for reading and quoting, SRT or VTT if the interview is going out as video with captions. In SRT the name is written in only where the speaker changes, rather than on every line; VTT carries it as a voice tag that costs no space on screen.

If you'd like to see how this works on one of your own recordings, you can try it without an account.