Speech recognition and models
Hallucination
Text the model produces without there being any basis for it in the audio.
A hallucination is text the system writes that was never said. Not a mishearing, but an invention: whole sentences that look plausible and that nobody uttered.
This is a real error type in modern speech recognition, and it is more dangerous than ordinary errors precisely because the result looks right.
Why it happens
The model predicts what is likely to come next, based on what it has heard so far. When the audio evidence thins out, it keeps predicting. With nothing to correct against, it produces the most likely text rather than nothing.
Where it most often occurs
- In silence. Long pauses are the classic trigger. The system fills the gap.
- In noise. Ventilation noise or music without speech can turn into text.
- At the end of recordings. Especially if the file ends abruptly.
- On very unclear speech. Where the audio barely carries, the model guesses more.
Typical symptoms are repetitions of the same sentence, or phrases that belong somewhere else entirely, for instance politeness formulas from the subtitle data the model was trained on.
How to catch it
- Proofread against the audio, not just against the text. Hallucinations are linguistically flawless, so they do not fail a read-through on their own.
- Look for repetitions. The same sentence two or three times in a row is nearly always this.
- Pay extra attention around pauses and in the last paragraph.
- Strip long silences before submitting the recording, if the tool does not do it for you. See voice activity detection.
For material that will be quoted, in research, journalism or legal cases, checking against the audio is not optional.