We systematically map the spectral vulnerability profile of OpenAI's Whisper speech recognition model by applying narrowband frequency notches (200–4000 Hz, 100 Hz steps) and measuring word error rate (WER) impact. For the word “glow”, 28 of 38 tested frequencies cause significant misrecognition, with characteristic hallucination patterns per band. Notching the /g/ burst zone (2000–2900 Hz) consistently deletes the onset consonant, producing “go”; notching the /l/ formant zone (1500–1800 Hz) deletes the lateral approximant. We generalize these findings across six words.
We introduce the dynamic anchor chain, a causal intervention that injects a trajectory-following sine wave into a notched critical band. This restores perfect recognition across 30/30 trials (0% WER) for “glow”.
Critically, we demonstrate a synthesis barrier: every approach to synthesize speech from extracted spectral features fails universally, producing empty or hallucinated transcripts. This barrier persists regardless of band count, amplitude balance, envelope structure, or phase coherence.