Computational Auditory Scene Analysis Based Perceptual and Neural Principles

摘要

A remarkable feat of the auditory system is its ability to disentangle the acoustic mixture and group the acoustic energy from the same event. This fundamental process of auditory perception is called auditory scene analysis. of particular importance in auditory scene analysis is the separation of speech from interfering sounds, or speech segregation. Consistent with specified objectives, this project made major advances along the following three directions. First, the problem of multipitch tracking was investigated in the context of multiple sound sources, and a robust algorithm for multipitch tracking of noisy speech was developed. The second advance is in monaural separation of voiced speech, where a new system was proposed that employs different strategies in the low- and the high-frequency range. A key element of the system is amplitude modulation analysis in the high-frequency range. Third, the problem of location-based separation was studied in the joint feature space of interaural time difference and interaural intensity difference, and a novel classification approach was introduced to optimally determine whether a target sound dominates in local time-frequency units. All of the three models were comprehensively evaluated and shown to be substantially superior to existing approaches.