Application Of Improved Spares Least Squares Support Vector Machine In Speech Recognition

Speech recognition is one of the most direct and convenient means of human-computer interaction, which is affiliated with the domain of multi-dimensional pattern recognition. Least squares support vector machine is a popular pattern recognition algorithm in the field of machine learning. As an extension of standard support vector machine, LSSVM possesses the advantages of small sample study, avoiding the "high-dimensional dimension disaster" and easy implementation of model training algorithm, making it suitable for complex speech signal recognition. But LSSVM has the drawback of solution sparseness missing, resulting in the improvement of model complexity and reduction of systems recognition speed. Therefore, research is carried out aiming at this problem in this paper, and the specific contents are as follows:(1) Based on the further study of the principles of speech recognition system and LSSVM, LSSVM is introduced into the speech recognition system, which overcomes such defects of traditional speech recognition method as requirement Hidden Markov Models for priori knowledge of the distribution and Artificial Neural Networks being prone to "over-learning". (2) After the careful study of the importance of the model parameters on the system learning ability and generalization ability, we propose using the scheme of combining particle swarm optimization algorithm with the ability of global optimization combined with K-fold cross-validation method to perform optimal parameter optimization, avoiding the problems of complex manual debugging and time-consuming automatic optimization of grid algorithm.(3) Based on deeply studying the reason of LSSVM sparseness loss and the influence of speech samples characteristic dimension on model performance, we propose an improved sparseness algorithm of least squares support vector machine based on independent component analysis. Firstly, independent component analysis is adopted for voice feature reduction; Then, after training the model, fast pruning algorithm is utilized based on independent component analysis for the nuclear matrix reduction, and the combination of the kurtosis and skewness is used as a measure of the importance of independent component to solve scheduling problems of independent component during the process of reduction. Experiments on Korean speech database show that the algorithm effectively implements model sparseness while ensuring the identification accuracy of the model.(4) Corresponding to the problems of the improvement of model complexity and the reduction of model recognition performance caused by the participation of non-support vector, this paper proposes an improved least squares support vector machine sparseness algorithm based on support vectors’ pre-selection, which takes into account two aspects including data mining and the geometric distribution meaning of support vector. Before training the model, the algorithm selects the union set of the key characterization samples extracted by K-means clustering algorithm and the boundary samples chosen by center distance ratio method as pre-support vectors, which effectively realizes the sparsification. Experiments on Korean and Aurora-2speech database show that this method well improves the recognition speed on the basis of almost no identification accuracy loss, reaching the goal of sparseness.