摘要
传统说话人识别框架大多建立在高斯混合模型(GMM)上的,然而这种浅层学习模型不能有效地表征数据特征之间的高阶相关性,识别效果较差。本文提出一种基于深度神经网络(Deep Neural Network,DNN)与基音周期(Pitch Period,PP)相结合的说话人识别方法,模型主线识别以对数梅尔滤波器组特征参数作为DNN的输入,通过训练DNN模型提取说话人的声纹特征;针对DNN模型阈值设定人的主观性影响,利用动态时间规整技术匹配说话人基音周期进行辅助识别。实验结果表明,这种双重识别方法等错误率可以达到1. 6%,较DNN系统与EM-GMM系统等错误率分别降低了1. 2%和2. 4%,并且在噪声环境中仍具有较好的鲁棒性。
Traditional speaker recognition frameworks are mostly based on the Gauss mixture model( GMM),but this shallow learning model can not effectively represent the high-order correlation between data features,thus the recognition effect is poor. In this paper,a speaker recognition method based on Deep Neural Network( DNN) and Pitch Period( PP) is proposed. The logarithmic Meier filter bank feature parameters are used as the input of DNN for mainline identification,and the voiceprint characteristics of the speaker are extracted through training DNN model. To eliminate the subjective influence of threshold setting in DNN model,dynamic time warping technology is used to match pitch period of the speaker for assistant recognition. The experimental results show that equal error rate( EER) of this dual recognition method reaches 1. 6%,which decreases respectively by 1. 2% and 2. 4% compared with DNN system and EM-GMM system,and this method still has good robustness in noise environment.
作者
张学祥
雷菊阳
ZHANG Xue-xiang;LEI Ju-yang(School of Mechanical and Automobile Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)
出处
《计算机与现代化》
2020年第1期122-126,共5页
Computer and Modernization
