Speech Separation Based On Computational Auditory Scene Analysis

Computational auditory scene analysis(CASA) is the study of auditory scene analysis by computational means based on psychoacoustics. Speech separation based on CASA is one of the principle study of speech signal processing. It is very important for speech recognition, multimedia retrieval and artificial intelligence.The most important issue for CASA is to choose the appropriate speech separation cues. Compared with voiced speech separation, unvoiced speech separation is significantly more challenging and little previous work has been addressed. According to the theory of auditory scene analysis, onset and offset corresponding to sudden intensity changes, can be applied to both voiced and unvoiced speech.Based on the principle of computational auditory scene analysis, this paper describes a model of speech separation by analyzing onset and offset of auditory events. The model first detectes onset and offset, and then generates segments by matching corresponding onset and offset fronts. Compared with using different speech separation cues, our system can separate both unvoiced and voiced speech. The model is evaluated with three kinds of corpus, and the evaluation shows that it can separate all the corpus with excellent performance and faster computing speed.