Cochannel Speech Separation Based On Computional Auditory Scene Analysis

With the development of information technology, speech separation is closely related with search engine and artificial intelligence, and speech separation based on computational auditory scene analysis(CASA) has broad application prospects in the field of multimedia retrieval and robotics research, which have gradually become research focus. Currently speech separation systems based on CASA are difficult to obtain satisfying results when handling cochannel separation.The reason is that most of the CASA system can not get accurate multiple pitch, thereby affecting the separation. And many separation systems use the training model in organization stage, which require the availability of pretrained speaker models and prior knowledge of participating speakers.In this paper, a cochannel speech separation method is proposed,which based on existing research. The main contents include:(1) A multipitch tracking method based on hidden Markov model(HMM) was proposed. Fisrtly, speech is decomposed into time-frequency units by peripheral processing mode. Then, the method calculates multiple pitch by hidden Markov model tracking algorithm based on Statistical properties, and label time-frequency units to obtain sumultaneous speech streams in the presence of more than one pitch. Experiments showed that this method was effective for calculating multiple pitch.(2)A sequence organization method based on clustering was proposed. Firstly, the method extract gammatone frequency cepstral coefficients from speech materials, and propose a objective function baed on between-and within-cluster scatter matrix. Then search for the optimal assignment of simultaneous speech streams by maximizing the between-and within-cluster scatter matrix ratio to separating the mixtures. Experiments showed that this method was effective for cochannel speech separation.