Reverberation Environment Of Isolated Word Speech Recognition Research

The robustness of speech recognition system is the key whether the technology ofspeech recognition can step into the stage of practical application. But the existence ofnoise or reverberation in real environment, especially in distant-talking speech recognitionsystem, reverberation will cause the change of amplitude, the delay of phase, the offset offormant and the production of other peaks in speech signals. What’s more, the tailingreverberant part also hides the weak energy sound of the following voice. It greatly reducesthe voice intelligibility, causing the decline of recognition rate. So it is important toovercome reverberation for realizing distant-talking speech recognition.This paper proposes one method combining the complex spectrum linear filteringwith U-GMM to improve the recognition rate of isolated word speech in the reverberationenvironment, which is on the basic research of reverberation basic characteristics andspeech recognition algorithm. The main research works included is as follows:This paper mainly studies the reverberation characteristic, against the features that thereverberation voice is added by many time orders and the pure signal is weaken. Accordingto this feature that the complex spectrum of pure speech signal in the complex spectrumdomain is commonly distributed near the origin, while the complex spectrum of roomimpulse response is far away from the origin. First, we take use of the discrete Fouriertransform the speech signal into the complex spectrum domain. Then, realize the reductionof reverberation through the blind deconvolution linear filtering the amplitude and phaseinformation of reverberation. At the same time, it will reduce the distortion of speech whilenot change the information of original speech.In order to solve the problem that the reverberation affection crosses multi-continuestime frame, this paper proposes a method to adaptively adjust the mean value andcovariance value of every frame in Gaussian mixture model (GMM) according to the prior vector of characteristics. Then take use of maximum expected algorithm parameters toachieve the optimal estimation. Combing the complex spectrum linear filter, the filteredreverberation speech is input into this model for matching recognition, in order to improvespeech recognition rate in reverberation condition.We conduct the experimental simulation using this method, taking use of subjectiveand objective evaluation to judge the reverberation speech recognition. The experimentalresults show that the recognition rate in this paper is obviously higher than other commonalgorithms, better removing reverberation and guaranteeing the minimum voice distortion.At last, this paper summary the research work, the shortcomings and deficiencies ofthe proposed algorithm and also explore the direction of future research.