Certified Robustness Significantly Reduces ASR Word Error Rates

Automatic Speech Recognition (ASR) systems are known to be vulnerable to both adversarial and benign audio perturbations, which can significantly degrade performance. Detecting these issues in deployed systems is challenging due to the lack of a "true transcription" oracle. Researchers have developed a novel certification-inspired mechanism to address this, significantly improving ASR robustness. This system employs a dual-gate diagnostic pipeline: a Two-Sided Atomic Audit that statistically certifies token existence and adversarial exclusion, and a Rank-Based Tournament for selecting the most accurate sequence. Evaluations across four diverse ASR architectures demonstrate impressive results, including up to a 55% relative reduction in Word Error Rate (WER). The mechanism also provides granular word- and sentence-level certifications, boosting recall and decreasing the correlation between confidence and WER, thereby enhancing overall acoustic security.