基于k均值和基于归一化类内方差的语音识别自适应聚类特征提取算法

发布时间：2018-03-01 15:13

本文关键词： 特征提取自适应聚类特征帧间独立假设归一化类内方差　出处：《清华大学学报(自然科学版)》2017年08期 　论文类型：期刊论文

【摘要】：语音识别模型中帧间独立假设在给模型计算带来简洁的同时,不可避免地降低了模型精度,增加了识别错误。该文旨在寻找一种既能满足帧间独立假设又能保持语音信息的特征。分别提出了基于k均值和基于归一化类内方差的语音识别自适应聚类特征提取算法,可以自适应地实现聚类特征流的提取。将该自适应特征分别应用在Gauss混合模型-隐Markov模型、基于段长分布的隐Markov模型和上下文相关的深度神经网络模型这3种语音识别模型中,与基线系统进行了实验对比。结果表明:采用基于归一化类内方差的自适应特征可以使得3种语言模型的识别错误率分别相对下降10.53%、5.17%和2.65%,展示了语音自适应聚类特征的良好性能。
[Abstract]:The assumption of inter-frame independence in speech recognition model not only brings simplicity to the calculation of the model, but also inevitably reduces the accuracy of the model. This paper aims to find a feature that can satisfy the inter-frame independent hypothesis and preserve speech information. An adaptive feature extraction algorithm for speech recognition based on k-means and normalized intra-class variance is proposed, respectively. It can be used to extract the clustering feature flow adaptively. The adaptive feature is applied to the three speech recognition models, namely, the Gauss hybrid model-hidden Markov model, the hidden Markov model based on segment length distribution and the context-dependent depth neural network model. The experimental results show that the recognition error rate of the three language models can be reduced by 10.535.17% and 2.65%, respectively, using the adaptive feature based on normalized intra-class variance. The results show that the speech adaptive clustering can be achieved. Good performance of features.
【作者单位】：清华大学电子工程系;
【基金】：国家自然科学基金面上项目(61374120)
【分类号】：TN912.34

【相似文献】