Semantic-Physically Conflicting Speech Perception and Human Cognitive Principle Inspired ASR System Design

Human exhibits great capacity in processing various kinds of complex and conflicting speeches in high interference acoustic scenes like cocktail party and many other public environments. Discovery of the auditory cognitive conflict control mechanism has great inspiration to automatic speech recognition (ASR) system design. The present research investigated the auditory cognitive conflict control mechanism in human brain through the event-related potential (ERP) technique under auditory perception conditions. Based on the wildly accepted Stroop paradigm, two types of auditory cognition experiments were designed to cover intra-channel and inter-channel conflicting speech perception. The ERP investigations indicated that a hierarchical conflict control mechanism plays an indispensable role in the both types of conflicting speech perception. A hierarchical feedback-controlled speech recognition system framework was thus proposed as an engineering model of the observations in this study. The framework offers a novel mechanism with multi-level feedback loops to modify the feature extraction algorithm and a multi-granularity feature fusion criteria to identify the speeches. This ASR architecture would greatly improve online speech recognition performance.