What the vendor numbers actually mean
Every speech recognition provider publishes an accuracy figure, and those figures look reassuring. The problem is not that they are false. The problem is what they were measured on.
- Advertised accuracy is typically measured on read speech — clean studio recordings, taken from public read-speech benchmarks.
- Independent measurements on spontaneous conversation give a markedly worse error rate for the very same models — roughly three to four times higher.
- A model fine-tuned on one specific language does measurably better on that language than a general-purpose one.
- In dialogue, with people talking over each other, even the fine-tuned model falls back again.
The two kinds of number do not contradict each other: they were measured on different material. The advertised figure comes from read, clean studio speech; the independent one from real conversation. Your recording is much closer to the second.
The gap between read and spontaneous speech
It is worth understanding why the difference is this big. Read speech is segmented, the sentences are finished, nobody talks over anybody, there are no filler sounds and no slips of the tongue. Spontaneous speech has all of that — plus speakers who abandon a sentence halfway, ask back, and talk at the same time.
The practical consequence is simple: the vendor's accuracy figure does not apply to your material. If you are transcribing a meeting, an interview or a podcast, count on an error rate several times the advertised one. That does not make the tool useless — you still save around an hour and a half of work per hour of audio — but it does mean that reviewing is not an optional step, it is part of the pipeline.
The hallucination that is born in the silence
There is a separate, nastier class of error: when the model does not mishear, but writes down something that was never said. One large study across more than thirteen thousand segments found this in roughly one and a half percent of segments.
The fresher data is much harsher than that. According to a measurement from September 2026, on non-speech audio — music, noise, silence, machine hum — the hallucination rate scatters between 21 and 76 percent, depending on the model. Which means the most dangerous places in the transcript are exactly the ones where nobody is speaking: the start of the recording, the pauses, the end. Those stretches are worth checking deliberately, because an invented sentence looks just as confident as a correct one.
Speaker identification
When there are several speakers, the machine tries to separate them. There is measured data on this too: the two systems examined worked at error rates around 15 and 17.5 percent.
In practice that means one sentence in six or seven ends up attached to the wrong name. For a meeting summary that is a serious problem, because "who committed to what" turns on exactly this. There are two ways to fix it: at recording time everybody speaks into a separate microphone, or during the review you pay separate attention to the speaker changes. The second is cheaper, the first is better.
The subtitle that has rules
A subtitle is not the same thing as a transcript. A transcript contains every word; a subtitle has to be readable, and there are standards for that.
- Line length: at most 42 characters per line.
- Number of lines: at most two on screen at once.
- Reading speed: for adult content, typically 17 to 20 characters per second depending on the language, lower for children's content — captions for deaf and hard-of-hearing viewers are more permissive.
- One subtitle on screen: roughly 5 to 7 seconds; a longer block has to be split.
Those four numbers are what make a subtitle look professional. Automatic subtitlers typically just chop the transcript up by time, and they do not keep these rules — the lines come out long, and they change too fast.
The pipeline
1. RECORDING: a separate microphone per speaker, where you can.
This does more for the transcript than any software will.
2. TRANSCRIPT with a machine tool. Supply a WORD LIST if the
provider allows one: company names, products, colleagues, jargon.
3. REVIEW, in three targeted passes:
- the silent and the musical stretches (it hallucinates here)
- the names, the numbers, the jargon
- the speaker changes
4. SUBTITLE from the corrected transcript, not the raw one.
Wrap at 42 characters, two lines maximum.
5. TIMING check: nothing faster than about 17-20 characters
per second, nothing on screen for longer than 7 seconds.
Common misconceptions
- "The vendor promises five percent error, so it is nearly perfect." That holds for read speech. On a real conversation the error rate is about four times higher.
- "Hallucination is rare, negligible." While someone is speaking, yes. In silence and over music the measurements say it can hit every second segment.
- "A transcript and a subtitle are the same thing, just timed." They are not: a subtitle has readability limits, and the automatic cut does not respect them.
- "It is bad in my language anyway, not worth it." It is worth it: even at a double-digit error rate, correcting is faster than typing. Just plan the review in.
What to take away
A machine transcript is a starting point, not a finished product — especially outside English, and with spontaneous speech. The three targeted checks (silence, names, speaker changes) catch most of the trouble, and for subtitles hold to the 42-character, two-line, 17-to-20-characters-per-second frame. If a multilingual version comes out of the transcript, Lipsync and Dubbing is the post that covers it; if you want to cut short clips out of it, Turning Long Video into Short Clips picks up from there.