Independent Component Analysis And Its Application In Preprocessing Of Speech Recognition

ICA is a higher-order statistics tool. From 1995, it has been more and more widely recognized and applied in diverse sub-fields of signal processing domain. Nowadays it is an important and indispensable research frontier in correlative fields of pattern recognition and signal processing. This paper is mainly engaged in the research of BSP theory based on ICA and its application in blind speech signal separation and speech enhancement for the preprocessing of speech recognition, attempting to secure some robust measurement of speech recognition.ICA theory is discussed in the paper, especially noted the Fast-ICA algorithms which has been widely recognized. Based on which, we did three aspects research as follows:This paper makes an attempt to apply these ICA algorithms in speech signal blind separation. For evaluating the performance of separating pure speech signal from mixed speech signal by ICA, we propose some subjective and some objective measures. In addition, we also implement an algorithm that can acquire pure speech signal from mixed speech signal by a single channel, traditionally ICA needing at least two signal channels.Upon comparing ICA with classical de-noising algorithms, we find out ICA precedes classical algorithms if intensive environmental noise presents, also ICA shows notable stabilization for vast SNR range. Although above results are driven from speech signal entangled with intense Gaussian whiten noise, their applicability also discussed in many other different noise environment. Furthermore, we propose a united de-noising algorithm based on ICA and wavelet. Tentative tests show that this united method holding more powerful speech enhancement capability.Finally, we implement a preliminary intelligent household conjunctive words speech recognition system, based on which we design a series of examinations. The results indicate that the preprocessing methods based ICA can enhance the recognition capability of the system.In conclusion,not only from the blind source separation of mixing speech, but also from the performance of speech enhancement or from the advance the recognition ratio of system, ICA is indeed an kind of efficiency methods for speech recognition preprocessing. The study of this paper will have positive meanings to speech analysis and recognition under noising environment, especially under intensive environment noise.