Automatic subtitles: what to check before you trust them

"99% accuracy" is the figure on nearly every automatic transcription site. It is a number that means nothing until you know what audio it was measured on — and it is almost always English, read aloud, into a professional microphone.

This article tries to be more useful: where automatic transcription actually goes wrong, why it goes wrong there, and how to check it on your own files in two minutes.

The number that matters isn't the percentage

The standard measure is word error rate: how many words are wrong per hundred. The problem is that it treats every error as equal, and for anyone publishing they are nothing of the sort.

A wrong preposition goes unnoticed. Your guest's name, mangled, is noticed by everyone — starting with the guest. Two transcripts with the same error rate can be one perfectly usable and the other embarrassing.

So when you evaluate a tool, don't look at how many errors it makes: look at where it makes them.

The four places transcription struggles

Proper names and brands

The worst case, because the model tries to map an unfamiliar sound onto a word it knows. An uncommon surname becomes something else that is plausible, and the transcript still reads smoothly — which makes the error harder to spot, not easier.

Acronyms and technical terms

Every field has its jargon, and speakers say it fast because they take it for granted. Acronyms read letter by letter fare better than the ones pronounced as words.

Overlapping voices

The hardest case of all, and harder than background noise — which is actually handled rather well. An interview where people talk over each other consistently produces worse results than a noisy recording with a single speaker.

⚠️ Worth knowing because it is counter-intuitive: the "bad" recording isn't the one in the noisy room, it is the one with two people talking.

Language switching inside a sentence

"I ran a check on the funnel performance." Spoken language in many industries is full of borrowed words, and the model has to work out on the fly that a given word should not be spelled in the surrounding language. It is the case where tools tuned for English do better, and badly configured ones do worse.

How to check it in two minutes, on your own files

You don't need a scientific test. You need a sample that represents your actual work:

1. Take two minutes of one of your real recordings — not your best file: your typical one.
2. Pick a stretch where you say proper names and at least one technical term from your field.
3. If you can, include a passage where somebody else cuts in.
4. Transcribe it with the tool and read it back **counting only the errors you would have to fix

before publishing**.

If there are two or three in two minutes, the tool is saving you time. If there are fifteen, it is moving your work from writing to correcting — which is duller.

How often we get it wrong, measured

The advice above applies to us too, so we followed it. We had echoflow transcribe an Italian podcast whose text we know word for word, because the audio was generated from a script: a 5-minute-25-second business interview with 28 figures and 11 proper names — three people, a company, two cities, two research institutions, a report, a book and the podcast's own name. The result, measured on September 25, 2026:

• 2.5% of words wrong (21 of 835), counting words changed, lost and added;
• 25 of 28 figures right;
• proper names right 24 of 25 times.

The three wrong figures say more than the percentage. "Forty-eight hours" becomes "40 hours", and "from four hundred and ten to two hundred and twenty kilometres" becomes "from 400 euros": smooth errors that nobody notices while reading, which is exactly the kind to look for. The missing name is the podcast's own, lost together with the opening line.

The measurement also changed our minds. For ten days we had used a transcription model that, tested on its own, heard figures better; measured this way, inside the real service, it dropped whole passages and in some runs added sentences nobody said. We went back to the previous model, the one behind these numbers.

What this measurement doesn't tell you. It is a single recording, with a single synthetic voice, clean and without noise: the easy case. There are no overlapping voices, which we called the hardest case above, and a real interview will do worse — by how much, this measurement doesn't know. And it isn't a comparison with other tools: it measures echoflow, not the market.

Something we built and then removed

It sounds sensible for a tool to warn you when it "isn't sure" about a passage, and to flag text below a certain confidence threshold for checking.

We built it, and we took it out. The reason: confidence scores measure how sure the model is, not how right it is. In our tests a real interview with overlapping voices scored worse than a pure-noise sample — the indicator couldn't tell legitimate speech from noise. An indicator that fails that way is worse than no indicator, because it teaches you to trust at exactly the wrong moment.

The practical conclusion is less elegant and more honest: the read-through doesn't go away. It shrinks, it concentrates on the right places, but it is not delegated to a traffic light.

What to do with subtitles once you have them

The SRT file is only the beginning. From the same timed text come the chapters, the quotes, the points to cut clips from: it is why it pays to treat the transcript as infrastructure rather than a final output — we cover it in Repurposing video.

---

If you want to try it on one of your own files, echoflow includes 14 days with no credit card. The advice is the one above: upload the typical file, not the best one. It is the only way to know how it will go in a normal week.