Research Of The Key Issues In The Support Vector Machine For The Anti-noisy Speech Recognition

Speech recognition is an important technology of human-computer interaction and pattern recognition. It has wide application prospect, great theoretical significance and practical value in speech recognition.Up till now most of the speech recognition systems are only adapt to recognize “clean" speeches. When there are background noises or the training and testing environments are not the same, the performance of the recognition system will decline sharply. Therefore, the speech recognition system performance needs to be improved; however, the common method of speech recognition can not achieve a satisfying result. As a new pattern of recognition method, the theory of Support Vector Machine (SVM) is based on the Structural Risk Minimization principle and the VC dimension theory. By using the Maximal Margin Principle, the kernel function methods, SVM can competently solve classification problems for small samples, the nonlinears, the high dimensions and the local optimal solution etc. It suits the characters of speech signals. And it has been applied in the field of speech recognition.The research of this paper is set according to the core content of how to improve the performance of speech recognition system based Support Vector Machine. This paper carries out several analyses and researches about application of Support Vector Machine in speech recognition from the following perspectives respectively: the selection of multi-class classification method for anti-noise speech recognition, selection and construction of kernel functions, the local Support Vector Machine algorithm for accelerating the training speed of the speech recognition system, this paper carried out several analyses and research. The main work and achievements of this paper can be concluded as follows:(1) By studying the theoretical basis and the basic principle of SVM in detail, theoretically analyzing the robustness of the Support Vector Machine deeply, the SVM is selected as the recognition method in this paper. Then it builds a speech recognition system based on support vector machine, analyzes the speech recognition's basic principle, the overall flow, the model training and the pattern matching etc in detail, studies the design and recording the flow of speech database, establishes the Chinese 500-word speech database.(2) In order to improve the performance of anti-noise in speech recognition,this paper studies thoroughly the strategy for solving the multi-class classification problem, introduces the existing M-ary and Error Correcting Output Codes (ECOC) theory in communication system into the application of Support Vector Machine in speech recognition for the first time. And the experiments prove that in both the clean environment and the noisy speech environment, the results of speech recognition by the ECOC illustrate an excellent robustness, better generalization ability and better performance ofanti-noise.(3) The kernel function is the core of the Support Vector Machine, and its selection directly affects the performance of SVM. Therefore, the selection and construction of kernel function plays a very prominent role in the theoretical study of Support Vector Machine. This paper presents two new kernel functions,which are the Logistic and the Optimal Relaxation Factor (ORF) kernel function.This paper theoretically proves that the Logistic and the ORF function are the Mercer kernel function respectively. The experiments also show that the new kernel function's effectiveness on the mapping trend, the bi-spiral test, and the two isolated words, which are Vowel and TIDigits, speech recognition problems.The paper draws the conclusion that the new kernel functions have better generalization ability and anti-noisy ability.(4) To shorten the real-time of the speech recognition system, and to accelerate training speed of the standard SVM, considering the characteristics of the local similarity of speech samples in features and the weak correlation of between non-adjacent samples, this paper presents an improved local SVM algorithm, gives description of the improved algorithms, the proof of the local kernel function and specific processes. The four experiments of the speech database, which are Vowel, CASIA digit string, ISOLET, Chinese 500-word,verify that improved local SVM algorithm is able to shorten the training time effectively in speech recognition.