The Experimental Study Of Blind Mixed-Speech Signal Separation

During these years, Blind Source Separation (BSS) has been a focus in the research field of signal processing. Because it can reconstruct the original signals from the observed mixture signals without enough prior knowledge of the mixing system and source signals, it is considered as a powerful tool in the field of wireless communication, radar, audio signals processing, medical signals processing and image enhancement.The BSS for audio mixtures is original intention of BSS techniques and also is full of difficulties and challenges in signal processing realm. In this thesis, the problems of instantaneous audio mixture separation are studied by experimental method. The main research works of this thesis are as follows:First, the mathematical description of BSS is given after a relatively thorough introduction to the historical perspective of BSS, including the mathematical model of BSS, the assumptions made about BSS problems and the uncertain of BSS. The algorithms used in this thesis are particularly introduced, which are: the Information Maximization algorithm (Infomax); the Fixed-point algorithm (FastICA); the Jointly Approximate Diagonalisation of Eigenmarices (JADE); the Non-negative Matrix Factorization (NMF).Then, the chapter describes the real-time audio mixtures acquisition system, which was designed in the thesis to get the signal data. The hardware of the system is composed of the USB-1208FS data acquisition card, microphones, powers etc. The software of the system is a data acquisition program; it is implemented by LabVIEW language.Finally, we use the algorithms which have introduced before to separate the audio mixtures that acquired from the system. By comparing experimental, we study some influencing factors for the estimated signals, which are the correlation of the frequency, the speech manner of the person and the distributing of the sources and the microphones. Also the appraisal of those algorithms would be given. The results of the experiments prove out that the FastICA algorithm has better effect in the audio mixtures separation. When the correlation of the frequency is small and the distances of the sources is long, the algorithms based on the ICA have better separation.The work and experience in this thesis has some referenced meaning in actual speech separation projects.