An audio-to-text transcription example shows what spoken words look like once they are written down: timestamps down the left, one block per speaker turn, and speaker labels wherever the voice changes. The two standard styles are clean verbatim, which drops filler words, and full verbatim, which keeps every "um" and false start.
That much every guide agrees on. What almost none of them do is let you hear the audio the sample came from.
That matters more than it sounds. A transcript of clear studio speech and a transcript of four people round a table with one microphone in the middle look nothing alike, and only one of them resembles the recording you actually have. So below are three real recordings — a meeting, an interview and a voicemail — each with the transcript our pipeline produced, following along word by word as it plays. Nothing is retyped or tidied.
For the meeting, the source corpus publishes a human reference transcript, so the last section puts our output next to it and gives the exact error rate.
Example 1: a meeting, three speakers, one distant microphone
Press play. The highlight follows the word being spoken; the speaker label changes when the voice does.
This is the hard case, deliberately. The microphone is not on anyone's lapel; it is sitting in the middle of the table catching everybody at a distance, plus the room. Three people are talking and they interrupt each other.
Things worth noticing, because they are what a real meeting transcript looks like:
That last point is the honest weakness of every automatic system on far-field audio, and the accuracy section below puts a number on it.
Example 2: an interview, close microphone
Same pipeline, completely different reading experience. One voice at a time, a microphone close to the speaker, no crosstalk. The sentences arrive whole, punctuation lands where the speaker paused, and there is nothing to untangle.
If you compare this with the meeting above, you have the single most useful piece of knowledge about transcription accuracy: the recording decides the transcript. Not the software. A lapel microphone or a headset will do more for your transcript than any setting in any tool.
Example 3: a voicemail, single speaker, sixteen seconds
Short single-speaker audio is the easiest case there is, and it shows the plainest transcript format: no speaker labels, because there is only one voice, and timestamps only where a new sentence group starts.
This is what "what does a transcript look like" means for most everyday files — a voice memo, a lecture recording, a dictated note. One column of timestamps, one column of text.
Clean verbatim vs full verbatim
Both styles describe the same audio. They differ in what they throw away.
Full verbatim keeps everything a person said, including the parts they would not want quoted:
```
0:08 I had to figure out how to how to do the same thing for the
summary, that's how I see it.
```
The doubled "how to how to" is not a transcription error. The speaker said it twice.
Clean verbatim removes the repetition, filler words and false starts, and changes nothing else:
```
0:08 I had to figure out how to do the same thing for the summary.
```
Which one you want depends entirely on what the transcript is for:
| Use | Style | Why || --- | --- | --- || Qualitative research, legal, medical | Full verbatim | Hesitation and repetition are data || Meeting notes, internal records | Clean verbatim | Nobody re-reads their own stumbles || Publication, subtitles, marketing | Clean verbatim, then edited | Readability wins || Language and speech research | Full verbatim, plus non-speech marks | `[pause]`, `[laughs]`, `[inaudible]` |A useful rule: transcribe full verbatim, then clean a copy. Going from full to clean takes minutes. Going the other way means listening to the whole recording again.
The multi-speaker format
When more than one person speaks, the transcript needs to say who. The standard layout is one block per turn, label on the left or above:
```
Speaker 1 So did you have to write it in the way that they wrote it,
is that what you did?
Speaker 2 Yeah, just search for specific string.
Speaker 1 Fish the data, I was just thinking that, yeah.
```
Three conventions worth following, all visible in the meeting example above:
1. Never merge two speakers into one block. A paragraph containing two voices cannot be quoted, coded or searched.
2. Label on change, not on every line. Repeating "Speaker 1" on six consecutive lines is what makes machine transcripts tiring.
3. Rename the labels once, at the end. Automatic systems output `Speaker 1`, `Speaker 2`. Replacing those with real names is a two-minute pass, and it should happen after the transcript is final, not during.
What a video transcript looks like
A video transcript is an audio transcript — the picture contributes nothing the text can use. The only real difference is what you do with the timestamps afterwards:
The same file can produce both, which is why exporting to SRT is a format change rather than a new transcription. If you need captions, transcribe first and cut second.
The accuracy nobody publishes
Here is where transcription sample pages stop being useful. SpeakWrite states "99%–100% accuracy". TranscribeMe states "99%+". Neither publishes the audio those numbers came from, or the reference transcript they were scored against — so the numbers cannot be checked, and they cannot be compared with each other.
The meeting above can be checked, because the AMI corpus ships a human reference transcript for it. Here is our output for that 48-second window against that reference:
| | Human reference | Our transcript || --- | --- | --- || Words | 208 | 163 || Utterances / segments | 20 | 10 || Speakers | 3 | 2 || Short reactions ("okay", "right", "mm-hmm") | 7 | 2 |Word error rate on this excerpt: 30.8% — 41 substitutions, 17 deletions, 6 insertions.
Across the whole 35-minute meeting there are 17 comparable 45-second windows. The median error rate over them is 36.8%; the best window is 15.5%.
Three things that number is telling you, and one it is not.
It is telling you that far-field meeting audio is genuinely hard. One microphone in the middle of a table, three people, overlapping speech: this is close to the worst realistic case, and it is why the number looks nothing like "99%".
It is telling you where the errors go. Look at the speaker row: three voices in the room, two detected. The third person in that meeting almost never says a full sentence — they say "okay", "yeah", "mm-hmm". Five of the seven short reactions in the window are missing entirely. Errors do not spread evenly across a transcript; they concentrate on short turns and quiet speakers.
It is telling you what to fix first. Nothing in the software closes a 30% gap on that recording. A microphone per speaker does.
It is not telling you what your file will score. Listen to the interview in example 2 — same pipeline, close microphone, and the transcript reads like prose. That is the ordinary case. The meeting is the stress test.
How to get a transcript of your own audio
1. Record as close to the mouth as you can. A £20 lapel microphone beats every software setting.
2. Give the file a language if you know it, rather than leaving detection to guess on the first few seconds.
3. Turn on speaker identification for anything with more than one voice, before transcribing rather than after.
4. Choose the verbatim style up front — full if the hesitations matter, clean if they do not.
5. Read the first minute against the audio. Names, jargon and acronyms are where errors cluster, and they are worth fixing by hand.
You can transcribe audio to text or convert an MP4 to a transcript directly, with speaker labels and timestamps, and export to SRT if you need captions.
FAQ
What does an audio transcript look like?
A column of timestamps and a column of text, one block per speaker turn, with a speaker label wherever the voice changes. Single-speaker audio drops the labels; multi-speaker audio keeps them.
What is the difference between clean and full verbatim?
Full verbatim keeps every word actually spoken, including "um", repetitions and false starts. Clean verbatim removes them without changing anything else. Research and legal work usually need full; meeting notes almost never do.
How accurate is automatic transcription?
It depends on the recording far more than on the software. On the far-field meeting above, our error rate against the published human reference is 30.8%. On close-microphone speech like the interview example, transcripts read cleanly with only occasional name and jargon errors. Treat any single accuracy figure quoted without the audio it came from as marketing.
Do transcripts include speaker names?
Automatic systems output generic labels — `Speaker 1`, `Speaker 2` — because nothing in the audio says who is who. Renaming them is a manual pass, best done once the transcript is final.
Can I get a transcript of a video?
Yes. Only the audio track is used, so a video transcript is produced exactly like an audio one, and the same file can be exported as captions.