Why does Whisper add text that was never spoken?

Because it fills silence and music with text that commonly accompanied quiet in its training data. The best known example is a subtitle community credit line appearing at the end of a clip that contains no speech, along with sign offs such as thanks for watching. A guard should remove known credit lines and treat sign offs as suspect when they are the whole transcript or arrive impossibly fast.

How do I set a threshold to catch transcription hallucinations?

Measure it on real speech instead of reasoning about it. My first rule, four words inside half a second, deleted 25 real words from a clip of fluent Hindi; the densest real five word window I measured was 0.52 seconds, so the shipped rule flags five words inside 0.35 seconds. Test every threshold against genuine fast speech in the languages you serve before it runs.

Can one signal reliably detect invented transcription text?

Not on its own. Speed, the model's no speech probability, its average confidence and text repetition are each wrong often enough to delete real words, so the safer design acts when two signals agree. Garbled song lyrics, for example, look like nonsense but are real, and a single signal guard would remove them.

Your Own Product Hallucinates Too

¶ WAR STORYTHE WORDS NOBODY SAID

01Captions for a silence

I spend most of this blog worrying about what AI models say about other people's products. This entry is about mine. A video tool I run turns speech into captions with a well known transcription model, and that model has a habit anyone who has used it will recognise. Give it a stretch of silence or a piece of music and it will sometimes write words that were never spoken, most famously a subtitle credit line thanking a volunteer community that had nothing to do with the video.

Why it invents

I will not pretend to know the internals, but the pattern is easy to observe. The model was trained on enormous amounts of subtitled video, and subtitled video tends to end with a credit line or a sign off. Faced with audio that contains no speech, it produces the text that most often accompanied quiet in what it learned from. A model fills silence with whatever usually came next in its training, and for video that is a credit line. On a captioned clip that line is not harmless. It is a false statement burned into someone else's video.

// THE CALIBRATION~~SECONDS TAKEN TO SPEAK FIVE WORDS~~

0.0s0.250.500.751.0s

FLAGGED AS INVENTED

0.35sthe rule that shipped

0.52sfastest real speech I measured

0.625smy first guess, scaled to five words

25real words the guessed rule deleted from one clip of fluent Hindi

9words a second reached by fluent Hindi in the test clip

5signals in the final guard, none of them trusted alone

The first guess was four words in half a second, which is 0.625 seconds for five. Real speech was faster than my limit for fake speech.

¶ WAR STORYTHE FIRST DRAFT

02The rule that deleted real speech

The obvious signal is speed. Invented text often arrives impossibly dense, many words packed into a fraction of a second, so the first draft of the guard flagged any four words spoken inside half a second. It seemed conservative. Nobody talks that fast.

Somebody talks that fast

I tested it on a real clip of fluent Hindi before shipping, mostly out of habit. The guard deleted 25 real words. The model splits Hindi into short particles, and a fluent speaker reached about nine words a second, so ordinary speech sailed straight through my limit for fake speech. I could have guessed the threshold. Guessing is how you ship a filter that removes real sentences. I then measured the densest genuine five word window I could find, which was 0.52 seconds, and set the rule at five words inside 0.35 seconds, with the margin on the side of real speech.

Lyrics are not noise

The same instinct saved a second case. A song transcribed with its title as a prompt came back as garbled lyrics, which look exactly like nonsense and are in fact real words, sung. Without the prompt, the same clip produced the credit line. Garbled is not the same as invented, and a guard that cannot tell them apart will delete the singer.

EXTRACTED — THE SENTENCE THIS ENTRY EXISTS FOR

A threshold is a claim about the world. Test it against the world before it runs.

¶ WAR STORYWHAT SHIPPED

03Five signals, none trusted alone

The guard as it runs today

The final guard runs inside the transcription step itself, so every path that produces captions passes through it, including the two that run at export. It removes known credit lines always. It removes sign offs only when they are the whole transcript or arrive impossibly fast. It flags five words inside 0.35 seconds. It uses the model's own confidence, but only when a high probability of no speech and a low average confidence agree, and it treats heavily repeated text as suspect. Any single signal is wrong often enough to hurt someone. Two that agree are usually right. One more trap nearly made all of it pointless: the detailed timing data the rules depend on has to be requested explicitly, and without that request it comes back empty, so every segment rule would have passed everything while looking perfectly healthy.

Why this belongs on a citation blog

Every argument on this blog about being quoted accurately runs in the other direction too. If a tool you run puts words into the world, it is a source, and it can be a wrong one. The discipline is the same one I ask of everyone else: numbers instead of adjectives, outcome data over instinct, and the raw output rather than the rendering. When an empty result has nowhere useful to go, the tool now offers to take the lyrics as pasted text instead of showing nothing, because a blank caption is honest and a confident invented one is not.

It matters more as software starts acting on its own output. A caption file is read by people, but it is also read by whatever summarises the video next, picks the best clip or writes the title. An invented sign off in a transcript becomes an invented line in a summary, and nothing downstream can see where it came from. The cheapest place to stop a false sentence is the moment it is first written.

Tomorrow: the pricing page a machine can quote, and why publishing the number is the whole strategy.