GhostPrefix investigates reusable audio prefixes that manipulate automatic speech recognition without knowing future speech or synchronizing with its waveform. The framework combines token-level and representation-level objectives to support semantic steering, targeted transcription, and transcription degradation. Evaluations cover 13 ASR models from four families in digital and over-the-air settings, with downstream effects examined in voice assistants and AI note-taking tools.
GhostPrefix is available as an arXiv preprint. Visit the project website for artifacts and demos.