STUDIES OF AN ADVANCED AUDITORY MODEL AND THE APPLICATION TO IMPROVE THE ROBUSTNESS OF CONTINUOUS SPEECH RECOGNITION

点击翻译按钮获取中文摘要

Our final goal is to develop a reliable continuous speech recognition system based on a model of human auditory system. So, we have studied as follows :(1) On the base of a subword-unit-based isolated word recognizer (VQ-SWR) with the discrete hidden Markov models (DHMMs) as a recognition tool, which we developed before, the research to improve the robustness for speakers and some environment noises have been done. As experimental results, findings can be summarized as follows :[1] A new recognizer with the DHMMs replaced with the semi-continuous HMMs have been developed. Experimental results showed a considerable improvement of the new recognizer in speakerindependency.[2] We have developed a new subword-unit-based isolated word recognizer incorporated a multiparty and a speaker adaptation step on the base of the VQ-SWR.This is made up of DHMMs and a learning vector quantizer (LVQ) incorporated a feedback of information on the classification of input subword which is obtained from the … More output of the LVQ.Experimental results showed that the new recognizer performance including the robustness for speaker and noise in stationary states is higher than those accomplished with the conventional recognizer VQ-SWR.(2) To aim at achieving higher word recognition rates and higher noise robustness than the VQ-SWR, we have proposed a new recognizer (CM-RN-SWR) made up of a model (NLF-COM) of human cochlea called "a nonlinear feedback model for cochlea", a simple multi-layer recurrent neural network (RNN) which has feedback connections of self-loop type, and DHMMs for words. The NLF-COM and the RNN which were developed before by us has been used as a model of the human auditory system, and as a kind of spectrum analyzer for speech sounds and a subword recognizer, respectively. Experimental results showed that recognition accuracies for clean speech and speech in the presence of pseud-white noise are considerably improved in speaker-dependent applications in comparison with the VQ-SWR. Less