Whisper hallucinates text on silence in 9.2% of transcripts

A developer audited 2,000 voice clips from the Preterview interview app and found Whisper large-v3 invented text in 183 of them (9.2%), mostly phrases like "Thanks for watching" over trailing silence. The built-in no_speech_threshold failed because confident hallucinations pass the logprob check; enabling Silero VAD with speech_pad_ms=400 and condition_on_previous_text=False cut the count from 183 to 4.

• 183 of 2,000 clips (9.2%) contained hallucinated Whisper text
• Clips with 3+ seconds of trailing silence hallucinated 23% of the time
• no_speech_threshold failed: 129 hallucinations had avg_logprob above -1.0
• Silero VAD and disabling condition_on_previous_text cut cases from 183 to 4