How Hackable Is Your Speech Quality Metric? A Corrected Protocol, a Benchmark, and What Patching Buys
The study introduces a corrected protocol to measure how hackable speech quality predictors are, fixing flaws in existing methods regarding perturbation chains and attacker variance. Benchmarking four predictors reveals significant differences in vulnerability. Auditing defense mechanisms shows that patching only hardens the predictor's training space, and post-trained enhancers can actually worsen performance against unpatched models.