Elimination Of Ambiguity For Blind Source Separation Of Mixed Speeches In Frequency-Domain

Speech signal separation has great research significance in some aspects such as speech recognition and computer auditory, and has become the focus of academic research. Blind source separation (BSS) estimates the source signals just from observed signals without any knowledge of source signals and transmission channel parameters. BSS has become a main method for speech signal separation since it does not require much prior knowledge.In most cases, the real mixed speeches are convolutive mixtures. Frequency-domain BSS methods can transform the time-domain convolution operations into frequency-domain multiplication with small computation and fast speed, thus has become a mainstream method for blind deconvolution. However, the inherent scaling and permutation ambiguities of BSS have great influence on the frequency-domain methods, especially the latter. If the ambiguities are not well solved for the frequency-domain algorithms, the speech separation will be largely degraded or even lead to the failure of separation.In order to eliminate the ambiguities of the frequency-domain BSS methods for convolutive speech mixtures, the main work of this thesis includes three aspects.(1) studying the principles of two kinds of scaling ambiguity elimination methods, the minimum distortion method and normalization method, and comparing the two methods’ effectiveness of eliminating the scaling ambiguity in BSS algorithms, such as JADE, KM-F, CMN by simulation experiments. The minimum distortion was demonstrated more effectively.(2) comparing and analyzing two kinds of different distance functions based on different prior information of source signals, and studying a semi-blind BSS algorithm which could eliminate the BSS permutation ambiguity in frequency-domain. The semi-blind algorithm based on energy distance function was verified more effective to eliminate the permutation ambiguity. Besides, the semi-blind BSS algorithm obtained better performance in comparison with the BSS algorithm reordering the signals after separation.(3) by studying several typical probability density distributions and the distribution characteristics of speech signals in different frequency bands, a sub-band independent vector analysis (IVA) algorithm was proposed, which used different distribution models in different frequency bands to better solve the permutation ambiguity. Extensive simulations with simulated and real speeches showed that the proposed IVA algorithm was more effective to eliminate permutation ambiguity than the original IVA algorithm.