An Interpretable Biomimetic Auditory Framework for Target Speech Enhancement via Pitch-Synchronous Delay Superposition

The human auditory system can selectively prioritize a target speaker in the presence of competing speech, providing a biological model for the design of interpretable auditory systems. In this study, we present a biomimetic computational auditory framework that translates temporal-processing principles of human hearing into an engineering mechanism for target-speech enhancement. Motivated by the distinct temporal characteristics of transient and quasi-periodic speech components, the proposed framework employs two complementary processing pathways. A first-order differential pathway emphasizes transient consonant cues, whereas a delay-superposition pathway exploits the quasi-periodicity of vowel signals to reinforce target-related periodic structures. The enhanced components are subsequently screened through an enhancement-based criterion and selectively reconstructed using zero-crossing-based signal replacement. Experiments on CRM two-talker speech mixtures demonstrated effective target-speech enhancement across multiple speaker-pair conditions, including both same- and cross-gender mixtures. The proposed method consistently improved target selectivity and preserved the principal spectral characteristics of the designated speaker. Psychoacoustic experiments were further used to examine the perceptual relevance of differentiated consonant-vowel processing and selective auditory attention. Rather than reproducing auditory neural circuitry directly, the proposed model provides an interpretable engineering abstraction of biological temporal-selection principles. These findings demonstrate how auditory mechanisms can be translated into a transparent biomimetic signal-processing architecture and provide a basis for future bio-inspired and neuromorphic auditory systems.