Two Microphone Blind Source Separation Algorithm Research

Early studied for blind source separation is mainly researched on microphone array, so there are many excellent algorithms such as ICA algorithm, based on signal sparsity of research and so on. With progressing of blind source separation research, it was found that people’s ears even more magical for identifying speech signals,therefore, more and more scholars plunge into this new field.In order to simulate the human ears’ functions of separation, people use two microphones to simulate humans’ ears. For the current study dual-microphone blind source separation has not made substantial progress, we can study the model with blind source separation of underdetermined model,the mianstream separation algorithm for solving this case is using of the sparsity of signals. In this paper, researched the blind source separation of dual-microphone, and mainly to do the following research areas:1.The focus of analysis and comparison of the four categories of the independence criterion of ICA, maximizing likelihood estimation criterion, minimizing mutual information criterion, information maximization criterion and maximization of non-Gaussian criterion.2.The combination of the ICA algorithm and binary time-frequency masking(T-F BM) which described above, positive microphone blind source separation algorithm is studied.Firstly, the recorded two channel speech signal using ICA algorithm preliminary separation, obtain the two separated signals. Secondly, the two signals are converted to the frequency domain using short time Fourier transform(STFT), then obtained binary time-frequency masking matrix by the separated signals. Since the time-frequency masking can cause the generation of musical noise, and finally, we use the cepstrum smoothing technique processing time-frequency masking,and applying the smoothing mask to separate the source signals. Simulation results show that the algorithm separated signals closer to the real speech signals, and the performance has also been improved.3.For underdetermined model of dual-microphone blind source separation technology, this paper proposes a dual-microphone underdetermined blind source separation algorithm which based on cepstrum smooth technique.First, the mixed speech signal inputted to input of ICA for separation, the separated signals are processed by adaptive nonlinear binary time-frequency masking, and repeat the two steps until all of the source signals were separated.The separated source signals by binary time-frequency masking combined masked for improving the quality of the output signals, and dual channel stereo effect of the separated the speech signals can still be well preserved.Finally, all of the source signals is separated by the cepstrum smooth processing, get the final separation of the signals.Experiments show that the performance of the algorithm is better than DUET algorithm and BLUES algorithm, SNR gain is also improved.