How accurate is AI transcription, and where does it fail?

Why a single accuracy percentage tells you very little

Speech recognition is usually scored as word error rate: the share of words inserted, deleted or substituted against a human reference. Published figures for Whisper on clean read speech are low, and they are measured on benchmark corpora of exactly that. Your recording is not that.

Two files with the same word error rate can be very differently useful. A transcript with errors spread evenly across common words reads fine and you fix it as you go. A transcript that got every instance of the client's name wrong has a lower error rate and is far worse, because the wrong word is the one the document is about.

So this page does not quote an accuracy figure for this tool, because none has been measured here. The speed has: 8.6x faster than realtime for transcription on a 31-minute recorded lecture, Chrome, 16GB Windows laptop, or 7.1x including decode and a first-run model download. Accuracy on real speech has not been measured, and a number invented for a page would be worth less than saying so.

The five failures worth checking for

These are not random. Each one comes from something specific about how the model works, which is why they are findable rather than a matter of luck.

• Proper nouns and names. The model predicts likely words, and your colleague's surname is not a likely word. Expect a plausible common word in its place, spelled confidently
• Acronyms and jargon. Letter sequences come out as words, and domain terms become the nearest everyday phrase. A technical meeting needs its terms checked specifically
• Numbers. Dates, prices and quantities are transcribed by sound, so a digit can land an order of magnitude out while reading perfectly fluently
• Overlapping speech. Two people talking at once produces one voice's words, or an average of both that neither said
• Repetition loops. Over music, applause or a long silence the decoder can get stuck emitting the same phrase for as long as the quiet lasts

Why it loops, and why that is the failure that reaches a video

The other four are wrong words in otherwise sound text. The loop is different in kind: it produces many seconds of confident nonsense, and it happens exactly where nobody is watching the transcript, because nobody is speaking. That is how it survives a read-through and ends up burned into the picture.

The cause is the decoder predicting each next word from the words it has already emitted. With no audio evidence pulling it anywhere, the most likely continuation of a phrase is often that phrase again, so it repeats until something interrupts it.

This tool looks for it directly rather than hoping you notice. Each line is scored for internal repetition and flagged above a threshold, which catches the stuck-phrase case while leaving a genuinely repeated line alone. It is the one check here aimed at a specific known failure rather than at a general style rule.

What the editor flags, and what it cannot

Alongside repetition, the audit flags lines that run faster than 17 characters a second, lines shorter than 0.8333333333333334 seconds, lines left hanging over silence, and lines that overlap. Those are mechanical properties of the timing, so they are reliable.

What it cannot flag is a wrong word. The usual approach is to show the model's confidence and let you review the shaky lines, and that is not available here: the transcription pipeline in transformers.js returns text and timestamps, with no scores or token probabilities. There is a field for confidence and the flag fires when something genuinely supplies it, which today nothing does.

So the editor narrows down where to look at the timing, and for wrong words you read against the list above. Searching for the names and terms you know were said is faster than reading top to bottom, and it catches the errors that matter most.

Does the bigger model help?

There are two sizes here. The smaller is a 45MB download, the larger 80MB, and the larger is meaningfully better on accents, crosstalk and technical vocabulary. It is slower and holds more memory, which on a long recording in a browser tab is a real cost rather than a theoretical one.

Worth the larger one for anything you will quote from, and the smaller for finding a moment in a recording you already sat through. Neither is a human transcriptionist. For a legal or medical record where a wrong number has consequences, the honest answer is that you want a person.