What Was That Again? Certified Robustness for Automatic Speech Recognition

Researchers have made advancements in Automatic Speech Recognition (ASR) systems, improving their robustness and accuracy. A new certification-inspired mechanism has been shown to decrease Word Error Rate (WER) and increase recall.

ASR systems are known to be sensitive to both adversarial and benign perturbations^[1]. A recent study demonstrated that employing a certification-inspired mechanism can significantly decrease WER and increase recall. This is achieved through a dual-gate diagnostic pipeline, providing granular word- and sentence-level certifications^[1]. The study also evaluated four diverse architectures, showing up to a 55% relative reduction in WER. Meanwhile, another study analyzed the dominant architectures in speech language modeling, the Transformer and the Conformer, using a probing framework called Architectural Fingerprinting^[2]. The analysis revealed divergent hierarchies between the two architectures. Conformers were found to implement a 'Categorize Early' strategy, resolving phoneme categories 29% earlier in depth and speaker gender by 16% depth^[2]. In contrast, Transformers 'Integrate Late,' deferring phoneme, accent, and duration encoding to deep layers (49-57%)^[2].