Noise-robust speech recognition based on missing data recognition

Abstract

For several decades, many researchers have studied and proposed algorithms for robust speech recognition so that a speech recognition system be utilized not only in laboratory environment but also in real noisy environment. In general, the robustness is defined as a characteristic of recognition systems that they are less sensitive to adverse conditions even when they are trained in clean ones. But many problems exists in robust speech recognition. So, it is very difficult to make the robust speech recognition system. The goal of our work is to make the robust speech recognition system. To do this, we use Two methodology. Firstly, we use CDHMM(Continuous Density Hidden Markov Model). Hidden Markov Models(HMMs) are stochastic automata used to represent the distribution of stochastic process. The HMMs are very useful and widely used in speech recognition. Because mathematical formalism defining the HMMs is well known and automatic procedures are available to train the models. Secondly we use a ``Marginalization`` among various methods to treat the missing data. A marginalization method is used for processing missing data since it can be implemented with low complexity when applied to recognition system. And we also use ``Bounded marginalization`` for the advance in speech recognition. The measure selected for the detection is the local signal to noise ratio (SNR). To detect the present features needs an estimate of the local SNR from the noisy features │Y(w,t)│ and from the noise features │N(w,t)│. This estimate of the local SNR is then used to divide features into present and missing by the use of thresholding. To test the validate of propose methods, we experiment the performance of speech recognition by AURORA2.0 database, which has been used widely in the field of speech recognition task. In this thesis, we use only four types of noise in it. In this study, we proposed a new framework using missing data technique for robust speech recognition and evaluate...