Certifiable Robustness for Automatic Speech Recognition

A certification-inspired dual-gate diagnostic pipeline can reduce the WER of automatic speech recognition systems, improve recall, and weaken the Spearman correlation between confidence and WER. The pipeline uses two-sided atomic audits to accumulate statistical wealth, certifies whether a token exists and excludes adversarial examples, and then selects the winning sequence through a rank-based tournament. On four different architectures, WER is relatively reduced by up to 55%, and it provides word-level and sentence-level certification to enhance acoustic security.