The Research Of Speech Segmentation Based On Oscillatory Neural Network

Speech segmentation is an important and basic task for machine audition. People always hope they can segregate speech from a variety of interfering sounds, however, that's really easy for humans' acoustic system. Just as the famous"cocktail party effect", we can follow the words of whose we are interested easily, no matter how noise the party is. However, it's really hard for the computer. Many researchers have reported a lot of methods to resolve this problem such as Blind Source Separation (BSS) and Auditory Scene Analysis (ASA).And there are two kinds of methods in ASA, they are CAS A and ASA.In this paper we propose a multistage neural model to separate the mixed speech. The most important part of the model is a two- layer oscillator network, and it performs the stream segregation depends on the oscillatory correlation. In this network, a stream is represented by a set of synchronized relaxation oscillators, and each of the oscillators represents an auditory feature, the desynchronized oscillators correspond to different streams. Before the oscillator network, there are auditory periphery and mid-level auditory representations. Our model has been evaluated using a copus, and the copus includes the voiced speech which are mixed with noise. We find that the SNR of the separate result of the mixed speech improved a lot.