Hearing What Is Unspoken: Language Model Priors for Acoustic Adversarial Attacks

The researchers propose the Semantic Gambit attack, which uses predicted context generated in real time by an LLM to enhance acoustic adversarial attacks on streaming ASR, increasing corpus-level word error rate to 35.6%, three times that of the current best-performing streaming attack. The method targets the bottleneck that real-time ASR, under strict time constraints, can only transcribe based on incomplete information, supplementing predicted context with a low-latency LLM. The authors state that this shows common low-latency LLM tools can be exploited to systematically disrupt real-time ASR pipelines.