Frequency-domain Algorithms For Blind Source Separation Of Convolution Mixtures Of Speeches

In real environment, speech signals are often interfered by other speeches or noises, and the signals picked up by microphones are convolution of the speeches and the interfering signals. Therefore, deconvolution of disturbed speech signals is an important task for speech enhancement. Blind source separation (BSS) recovers the various independent components for source signals from observed signals without knowledge of source signals and transmission channel parameters. Because it requires little priori knowledge, BSS becomes a widely used method for speech enhancement.Algorithms of BSS for separating convolutive speeches can be divided into time-domain and frequency-domain methods.When the reverberation time is long, the time-domain method requires many parameters to learn, which results in long computing time and converge difficulty.In contrast, the frequency-domain method transforms the time-frequency convolution into product operation in multiple frequency bands, thus has simple computation and fast speed. As such, it has become a mainstream method for blind deconvolution. However, the frequency-domain BSS algorithm must solve the inherent permutation ambiguity which otherwise affects the separation performance, this is the key point of this thesis.The main work of this thesis includes three aspects: (1) study the existing order adjustment methods and examine the characteristics of speeches, a new order adjustment method based on energy correlation is proposed. It is applied to JADE, KM, c-FastICA, and CMN to separate the convolutive mixtures from simulated and actual speeches, the results of analysis and comparison of separation performance verify the validity of the method. (2) In the framework of CMN and constraint optimization, the energy correlation characteristics of speech is directly incorporated into BSS process, two new semi-blind CMN algorithms are proposed including gradient algorithm and Newton algorithm, and the permutation problem is solved together with BSS. Extensive experiments with simulated and actual speeches show that performance of the constraint CMN algorithms is higher than that of other BSS algorithm including CMN. (3) explore some influence factors for the frequency-domain method. The thesis carries out simulations with different kinds of window functions, analyzes the short-time stationarity of speech signals and the effect of different window functions; explores the relationship between the length of FFT and the order of the impulse response through extensive experiments. These results provide useful references for further study of frequency-domain algorithms.