大词汇连续汉语语音的mlp声学特征的研究.docVIP

大词汇连续汉语语音的MLP声学特征的研究 大词汇连续汉语语音的MLP声学特征的研究 　　摘要:短时声学特征参数如MFCC,PLP作为输入向量的高斯混合模型(GMM)的隐马尔可夫模型(HMM)的经典模型在大词汇连续语音识别系统(LVCSR)已取得了良好识别效果。但针对短时声学特征区分性差的特点,本文提出采用神经网络多层感知器(MLP)产生的两种类型差异特征HATs与TANDEM代替短时特征,分别训练GMM参数模型。实验结果表明,差异特征的GMHMM的LVCSR系统优于传统的短时特征的系统;为了更进一步提高系统识别率,该文又将两种类型差异特征HATs与TANDEM进行复合,构成MLPs特征流重建GMHMM,系统的错字率(CER)有2%～3.8%的明显改善。 　　关键词:多层感知器;差异特征;隐马尔可夫;高斯混合模型 　　中图分类号:TN912文献标识码:A文章编号:1009-3044(2010)13-3470-02 　　MLP Features for Large Vocabulary Continuous Mandarin Speech Recognition System 　　LV Dan-ju1, Ch. Plahl2, B.HoffmEister2 　　(1.Computer Science Dept., Southwest Forestry University, Kunming 650224, China; 2.Lehrstuhl Fur Informatik 6-Computer Science Department RWTH Aachen University, Aachen 52056, Germany) 　　Abstract: Typically Hidden Markov Model (HMM) in large vocabulary continuous speech recognition system (LVCSR),using short-term acoustic features vectors (MFCC/PLP) as input features to the Gaussian mixture model (GMM), has achieved good recognition results. However, for the poor of these short-term features on discrimination, neural network multilayer perceptron (MLP) are used to produce two types of discriminative features HATs and TANDEM instead of short features, and respectively, the corresponding GMM parameter models are trained. Experimental results show that the GMHMM the LVCSR system based on discriminative features is superior to the system traditional based on the short-term features; To further improve the system recognition rate, the two types of discriminative features HATs and TANDEM are combined as MLPs feature flow to retrain GMHMM, that leads to an absolute reduction of the character error rate (CER) of about 2%~3.8%. 　　Key words: MLP; discriminative features; HMM; GMM 　　对语音信号特征参数的研究是建立良好的语音识别系统的基础与关键。在过去的研究中,语音识别系统的特征提取成分主要包括频谱包络预测,特别是经过某些简单变化后的特征,目前前端大部分是基于短时轨迹(约10ms)信号分析的美尔倒谱(MFC)或是感知线性预测(PLP)。但这些传统的短时特征参数存在着对信号变化过于敏感,不能反映连续帧之间的相关特性,区分性差等方面的不足。近年来,国外很多语音研究机构在语音信号的特征提取、声学建模方面引入了神经网络ANN,其中由Berkeley国际计算机学院(ICSI)提出用基于MLPs的特征取代传统特征,系统的识别率得到