Analysis And Implementation On The Acoustic Model Based On Svm In Speech Recognition

Speech recognition system is divided into two major components, one is the acoustic model; the other is the language model. Speech acoustic model is a description of the model, the state of voice recognition by the state to describe the voice output to determine a voice; language model is the model output for each acoustic voice to determine, calculate each connection into the language of voice probability, to determine whether it can be a language.The major part of our understanding of acoustic models, acoustic models are based speech recognition system, a very important role.In most current speech recognition systems, acoustic model typically use hidden Markov model (HMM), the divided state, jump matrix, and the state output probability, the state of HMM Gaussian mixture model output largely taken (GMM) to achieve. This model uses a number of Gaussian probability density function based on the formation of the model results, the sample can make the complex simple, but the model can only describe a single probability samples of this class, the lack of correlation with other samples, which makes the Gaussian mixture model (GMM) relatively uniform, the promotion of the poor.GMM-based limitations, we will support vector machine in this model the application of acoustic research, support vector machine (Support Vector Machine, SVM) can make use of a limited number of samples measured on the unknown or the data can not be the optimal solution. When their training is not only considered the model belongs to this set of samples, but also took into account does not belong to this set of samples.Training a SVM requires the solution of a very large quadratic programming (QP) optimization problem,SMO breaks this large QP problem into a series of smallest possible QP problems. And it make the large scale samples SVM training be true . if the number of the training data is very large, we will use a filter to discard the train data of no use, then train the SVM model on the remaining vector data. Greatly improve the efficiency of the trainingFor many types of discrimination senone characteristics, we used in this method is to use multiple SVM corresponding to each of the two types of senone ruling +1 corresponding to the marker model, do not correspond to -1, the highest test scores senone selected . Accordingly, the data imbalance and posterior probability is also given solution.By experiments, it is proved that the SVM model can classify the senone well and the effect is better then GMM. Particularly the use of the improved SMO algorithm, with the sample size increases, the time complexity of the optimal solution obtained was significantly lower than the SMO Algorithm and Direct Solution for optimization.