Study On Array Speech Recognition Using HTK Toolkit

Automatic speech recognition system can obtain good recognition performance when thespeech signals are captured by a body-mounted microphone in a quiet environment. However,speech recognition performance of a single microphone degrades because the speech signalscan be seriously distorted by noise and reverberation. Especially with the distance betweenthe user and the microphone increasing, speech signal becomes significantly deteriorated,which leads to degradation of speech recognition performance. Thus, the use of microphonearray has been used as a front-end of a speech recognition system to improve the quality ofspeech signals, which can effectively improve the recognition performance of a distant speechrecognition system.This paper first describes the basic theory of speech recognition such as extraction offeature parameters, the acoustic model and language model and so on. Then, it also studies themethod of microphone array speech recognition based on parameter optimizationsystematically. An improved experiment platform design and realization method ofmicrophone array speech recognition are presented subsequently.Microphone array speech recognition is performed in two stages: array processing andspeech recognition. Currently, these two stages are performed separately. Namely, firstmicrophone array processing, then speech recognition. A parameter optimization-basedmicrophone array speech recognition method is adopted in the paper, and it can enhance theimportant components which are good for recognition, thus it improves the performance ofrecognition system effectively. It combines stages of array processing and recognition. HTKis used as the basic experiment platform. Through the analysis of the characteristics of theunique Chinese language pronunciation, HTK is modified to train HMM models and languagemodels. With the result that an improved continuous Chinese speech recognition system usingHTK toolkits is built for the distant speech recognition. In order to simplify the complicatedDOS steps of using single HTK modules, a series of experiments are done by the combinationof MATLAB and HTK in the paper. Simulation results demonstrate that better performancesof microphone array speech recognition system based on array parameter optimizationmethod are obtained in the lower SNR environment. Simultaneously the experimentprocedure is simplified.