语音识别中基于支持向量机的声学模型研究

在语音识别中,声学模型常常采用聚类后的状态（senone）作为建模单元,其输出的分布大都采用高斯混合模型（GMM）,该模型采用多个高斯分布的加权和,描述复杂的概率分布。然而,由于该模型采用最大似然估计的方法,仅仅考虑了本类样本的概率最大化,而并没有考虑与其他类之间的关系。本文尝试对基于支持向量机的声学模型进行研究,支持向量机（Support Vector Machine,SVM）方法是建立在统计学习理论的VC维理论和结构风险最小原理基础上的机器学习方法在利用SVM判别senone类别时,本文对每个senone建立一个一对多模型。然后把SVM模型输出结果中的距离软化为得分,选取得分最高者判决为此senone所属的类别实验证明,利用SVM能够很好地判别senone,区分性要优于GMM。

A STUDY ON THE ACOUSTIC MODEL BASED ON SVM IN SPEECH RECOGNITION

In most current speech recognition systems, the acoustic modeling components of the output of the recognizer are almost exclusively based on Gaussian Mixture Model （GMM）.but GMM is poor at the generalization performance. Support Vector Machine （SVM） is developed from the VC dimension Theory and Structural Risk Minimization Theory of Statistical Learning Theory（STL）. It does a trade--off between the quality of the approximation of the given data and the complexity of the approximating function to achieve better generalization. The generalization performance of SVM is better then GMM.When use SVM models to classify the senone specie, we should first construct a one--vs--rest model for each senone, and then score the sample to be classified, and choose the highest score as the specie of the sample.By experiments, it is proved that the SVM model can classify the senone well and the effect is better then GMM.