| Blind source separation (BSS) is a technique to recover the sources when only the observed signals are known. In this case, the sources and the transmission channel are unknown. Since few priori information is required in BSS, it has been widely used in speech signal processing. There are two kinds of models for the mixtures of speech signals. One is the model of instantaneous mixtures, the other is the model of convolutive mixtures. However, interference of the reflection and delay exists in the actual environment. Based on it, the model of convolutive mixtures is more suitable in the actual word. So BSS for convolutive mixtures of speech signals becomes more and more popular.The convolutive BSS algorithms can be divided into two classes:BSS in time domain and BSS in frequency domain. The frequency-domain methods can change the complex convolution operation into simple linear operation. Then the instantaneous ICA algorithms can be introduced into the frequency-domain BSS. Compared with the time-domain methods, the frequency-domain methods are more attractive because of the less calculating amounts and faster convergence. However, the frequency-domain algorithms have two ambiguous problems:amplitude ambiguity and permutation ambiguity. The problem of permutation ambiguity makes a very significant impact on the separation performance. In addition, noise interference can not be ignored in actual environment. In this case, we can not only consider the convolutive BSS, but also consider the noise suppression. So for the mixture signals in actual word, how to reduce the noise and achieve separation, and how to design an effective separation system, are the hot and difficult points of the scholars’research.For the above problems, we mainly research the frequency-domain algorithms for convolutive BSS. Especially for the permutation ambiguity, a novel permutation algorithm is proposed. In order to verify the effectiveness of the algorithm, and the further application to real environment, we propose the BSS system scheme in real world.real-world, and design a demonstration system. The contribution of this paper:(1) Research on three frequency-domain BSS algorithms:the complex value of the FastICA, fast kurtosis maximization algorithm (FKMA), and joint approximate diagonalization algorithm (JADE). The simulation results show the performance comparisons of the three algorithms. From the results, we analyze the problems of these algorithms.(2) Propose a novel permutation algorithm to solve the problem of permutation ambiguity. The correlation of the adjacent frequency bins is combined with the DOA method to make sure the robustness and accuracy. The algorithm can be summarized as the following three stages. In the first stage, we choose and permute the benchmark frequency bins. Secondly, permute the remaining frequency bins according to the permuted bins in the first stage. In the third stage, the frequency bins permuted in error are marked, and permute them by DOA estimation based on separation matrix as a supplement to alignment in stage2. Simulation results show that the proposed approach can achieve higher performance and it is robustness even in the real world.(3) Apply the frequency-domain BSS algorithm to the actual environment. And design a separation system for the real speech signals. Through the study of several common denoising algorithm, a novel denoising algorithm based on nonspeech auto-detection is proposed, which is more suitable for practical application. The system scheme is proposed, and the denoisng algorithm can be employed as the post-processing method in the separation system. Based on the system scheme, a system demo is also designed for BSS in real environment. By simulating the real environment, the experiments verify the effectiveness of system solutions. On the other hand, the experiments for real acquisition speech signals by the system demo show the separation is achieved successfully. Thus, the system has broad application prospects. |
PDF Full Text Request