Hallucination (ASR)
In speech recognition, a hallucination is transcript text the model outputs for words that were never spoken — a known Whisper failure mode on silence, music, or background noise.
A hallucination in automatic speech recognition is when the model produces confident-looking text for audio that contains no matching speech. It is a well-documented failure mode of Whisper-class models, which can emit repeated phrases, invented sentences, or leftover training artifacts during long silences, music, or noisy stretches rather than returning nothing.
This matters because a hallucinated line reads like a real quote but corresponds to no spoken words, so it corrupts a transcription even when the overall word error rate (WER) looks low. The common mitigation is to run voice activity detection (VAD) first, so the recogniser only decodes spans that actually contain speech; conditioning on cleaner context and disabling text carried over between segments also help. It is one reason two tools can diverge sharply on the same hard recording.