Whisper Hallucinations on Silence

A speech recognizer is asked to map audio to text, but not every segment contains words. The original Whisper paper describes several failure modes of sequence-to-sequence transcription, including repetitions, missed segment edges and hallucinations in which output text is unrelated to the audio. Long pauses, music or low-level noise can create conditions where the decoder produces plausible language despite weak speech evidence. The exact triggers vary with model, audio and decoding setup; silence does not always cause hallucination, and real faint speech must not be discarded casually. Whisper’s open-source transcription implementation has a no-speech probability and decoding thresholds to consider a segment silent. It also exposes a possible-hallucination silence threshold in a word-timestamp workflow. These are heuristics, not proof of what someone said. Tuning a threshold too aggressively can remove quiet words; leaving it too permissive can preserve invented text. A separate voice-activity detector may help segment audio, but it can also make mistakes on whispers, accents, laughter or distant speakers. Detection needs direct evidence. Compare the transcript with the recording at the reported time, look for words in regions without speech energy, and examine repetitions or abrupt topic changes. Human listeners may also struggle with noisy audio, so mark uncertain spans rather than guessing. Test negative examples containing silence and non-speech sounds, and positive examples containing faint real speech. Report false text and missed speech separately. A low average word error rate on spoken clips cannot establish safety on quiet segments that were not included in the test. This matters wherever a transcript becomes a record. An invented sentence can distort an interview, subtitle or care note even if the rest is accurate. Preserve audio, timestamps, model version and processing settings for audit. Do not use unsupported segments for decisions or publication without review. The correct fallback for insufficient audio evidence is uncertainty, not a fluent completion.