Research On Blind Source Separation In Frequency Domain For Convolutively Mixed Speech Signals

In recent years, blind source separation (BSS) of speech signals has become a hot topic in the field of blind signal processing, and it has widely potential application in speech recognition, mobile communication and remote meetings. There are two main methods to achieve BSS: independent component analysis (ICA) and time-frequency masking (TF Masking). BSS for instantaneously mixed signals has already achieved good performance. But in real environment, speech signals are mixed convolutively due to the reverberation and various reflection, which has made BSS quite difficult. Presently, more and more researches implement blind separation of convolutively mixed speech signals in frequency domain, in which the convolutively mixed sources in the time domain are transformed into the instantaneously mixed sources. Therefore, the well-developed methods based on ICA of instantaneous mixtures can be applied to separate signals in frequency domain. Compared with TF Masking, frequency-domain independent component analysis(FDICA) leads to small musical noise and spectrum distortion, however it converges slowly and will lead to permutation and scaling ambiguity. In this thesis, we present a novel method that introduces TF Masking method into FDICA to separate the convolutive mixtures. It aims to achieve better separation performance and significantly reduce computational costs, whilst the ambiguity problem of the frequency domain can be solved. Furthermore, by setting up the multi-channel audio recording platform, we conducted the experiment of block-online blind separation with a new post-processing method that optimized the ICA results, to improve separation performance with little additional computation. The set-up platform acts as the basis of our BSS system. This thesis is organized as follows:Firstly,Chapter 2 introduces the signal mixing models and microphone array models of BSS for deeper understanding. Then room acoustic impulse response with MLS(maximum length sequence) method is discussed. Chapter 3 gives comprehensive introduction of ICA and frequency-domain BSS of convolutively mixed speech signals, and then the FDICA algorithm based on Infomax theory is given.Secondly, at the beginning of chapter 4, there is a review of the separating methods based on time-frequency analysis of speech. Then, a novel method, which is based on FDICA, is proposed by combing time-frequency analysis of speech signals, it contributes to fast convergence, and significantly improves computational efficiency with good separation performance.At last, in chapter 5 we discuss how to set up the multi-channel audio recording platform by using ASIO driver, which will benefit our further research and construction of BSS system. After analyzing the existed problems of block-online blind source separation, a post-processing method using time-frequency masking is proposed. Experimental results indicate that the post-processing method improves separation performance and the whole separation algorithm shows good performance in separation and computational efficiency. The work in this chapter lays a solid foundation for the blind source separation system in real application.