基于多分类器融合的语音识别方法研究

Research on multiple-classifiers fusion for speech recognition

针对语音识别性能提高的问题,提出了一种基于多分类器融合的语音识别方法,该方法使用支持向量机(support vector machine,SVM),RBF神经网络与贝叶斯网络作为成员分类器,根据样本库中抽取的校验集计算各成员分类器的权值,以加权评分的投票策略进行决策融合。实验结果表明,通过多分类器融合的识别结果明显优于单个分类器,该方法是一种有效的语音识别方法,提高了语音识别系统的性能。

Fusion of multiple classifiers can be integrated superiority of the each classifier, and better recognition effect can be achieved. At present, the multi-classifier fusion is a hot topic in pattern recognition. In a speech recognition system, design of the classifier is the key point to the superiority. In this paper, a novel speech recognition approach based on multi-classifier fusion is proposed. SVM, RBF network and bayes net are fused by weighted voting strategy, and the weight is calculated according to the validation sets from sample database. The experiment results show that the performance of multiple-classifiers fusion is better than single classifier, and the method proposed in this paper is effective.