Research On Sub-band Blind Separation Of Mixed Speeches

In real environment, speech signals are often interfered by other speeches or noises, such as multi-speaker interference and environmental noises, and the multi-speaker interference is difficult to eliminate. Blind source separation (BSS) is a typical multi-signal separation method. BSS of speeches can be divided into time-domain and frequency-domain algorithms. The time-domain algorithm requires a large amount of computation and is difficult to converge. In comparison, the algorithm of blind source separation in frequency domain has the advantage as it changes the convolution to product, has and thus reduced computational complexity and improved convergence. Therefore, this thesis studies a sub-band separation algorithm for frequency-domain BSS.The main work of this thesis includes four aspects:(1) By using many speech data online, we transform speech into frequency domain by short time Fourier transform (STFT), analyze the kurtosis characteristic of many speeches and draw the distribution figure. (2) On the basis of the above analysis, we discuss the basic characteristic of the sub-band speech, and present a sun-band separation method. (3) According to the sun-band separation strategy, we design and adjust the nonlinear function for complex-valued negative entropy maximization algorithm, and perform separation experiments on mixed speeches. The results show that the sub-band method with changed nonlinearities can effectively improve the performance compared to the full-band approach using only one nonlinear function. (4) We describe several features of different window functions and compare the impact of different windows on the performance.