Speech-to-text hallucinations: how AI invents words in audio
aieducation
How speech models can add sentences that were never spoken, when it is most likely, and which safeguards reduce the risk.
Speech-to-text systems are measured by word error rate, but that number hides an important distinction. A misheard word is one kind of error. An invented sentence is another.
What a hallucination looks like
In silence, background noise, or unclear speech, a model may output a fluent, plausible phrase that does not appear in the audio. Because it reads naturally, it is easy to miss.
What the numbers say
- Whisper large-v2 gets about 4 words wrong per 100.
- The best current models get about 2 wrong per 100.
- In a June 2026 study, all seven modern models still invented words on hard clips.
Lower error rates are real progress, but they are averages. They do not guarantee that an individual transcript has no invented content.
Where the risk rises
- Long pauses and silence.
- Overlapping speakers.
- Heavy accents, specialist vocabulary, or poor microphones.
- Very short or clipped segments.
Safeguards
- Improve audio capture where you can.
- Choose tools that mark low-confidence words and link each line to its timestamp.
- Review high-stakes passages against the recording.
- Add automated checks, as described in validation layers for enterprise AI.
Responsible use
Where transcripts feed decisions about people, build in oversight. Our Responsible AI-by-Design Framework describes how.