Hearing the Unspoken: Language Model Priors for Acoustic Adversarial Attacks

2026-06-04

AI Summary

This work addresses the challenge of conducting effective adversarial attacks on real-time automatic speech recognition (ASR) systems, which are inherently constrained by causality and thus limited in their ability to leverage future context. To overcome this limitation, the authors propose Semantic Gambit, a novel approach that integrates a low-latency large language model (LLM) as a semantic prior into the acoustic adversarial attack framework. By dynamically generating semantically guided perturbations during live audio streaming, the method circumvents the causal information bottleneck of real-time ASR. Experiments demonstrate that Semantic Gambit substantially enhances attack efficacy, increasing the word error rate to 35.6% on standard benchmarks—three times higher than the best existing methods—thereby underscoring the critical role of semantic priors in real-time adversarial attacks against ASR systems.

Abstract

Automatic Speech Recognition (ASR) systems operating in real-time settings must process acoustic input under strict temporal constraints, where transcription decisions are inherently made on incomplete information. This causal constraint serves as an information bottleneck on attackers, significantly limiting attack performance. Our new Semantic Gambit attack breaks this causal limitation by augmenting the adversary with predictive context derived from a Large Language Model in real-time. Our experiments show that this form of augmentation can elevate the corpus-level Word Error Rate to 35.6% -- a three-fold increase over the current state-of-the-art. Ultimately, this work reveals how common, low-latency LLM tooling can be exploited to systematically subvert real-time ASR pipelines.

Problem

• Automatic Speech Recognition
• acoustic adversarial attacks
• causal constraint
• real-time ASR
• adversarial robustness

Innovation

• acoustic adversarial attacks
• Large Language Model (LLM)
• real-time ASR
• causal constraint
• Semantic Gambit