The Speech Recognition Based On Error-Correcting Output Code And Support Vector Machine

For meeting the increased demand of human-computer interaction technology, Speech Recognition has become a hot spot and has wide application prospect in machine learning. Having many advantages for solving multi-class classification, the Support Vector Machine is therefore an important tool of Speech Recognition. This paper describes the speech recognition process and some traditional identification methods. There are many development bottlenecks in these methods. As a new method, the Support Vector Machine is based on statistical theory. The training task involves optimization of a convex cost function:there are no false local minima to complicate the learning process. The approach has many other benefits, for example, the model constructed has an explicit dependence on the most informative patterns in the data (the support vectors), and hence interpretation is straightforward. However, Support Vector Machine is used to solve binary classification problem, how to solve multi-class classification, the researchers have made a number of algorithms:"one against one "method, "one against rest "method, DAGmethod and the " Error-Correcting Output Code(ECOC) "algorithm.This paper mainly studies the Speech Recognition system based on Error-Correcting Output Code and Support Vector Machine. The main achievements are as follows:(1) Further study the Error-Correcting Code algorithm. Its biggest feature is the classification error can be amended to a certain extent, so the system can give more accurate identification results.(2) In this paper we combined Error Correcting Output Codes with SVM to construct a Speech Recognition system. There are some different kinds of Error-Correcting Output Codes and we applied them in speech recognition.(3) One of the ECOC is Hadamard ECOC.Compared with other methods,Hadamard ECOC had much better recognition performance. But it has a drawback that when the number of multi-class is big, the system will cost more time to train the recognition network. To solve this problem, this paper proposed a sparse error correcting code of the Hadamard ECOC:1/2 and 1/3 sparse of Hadamard ECOC.In Speech Recognition system, the experiment results were good.