基于自动编码器的语音音色客观评价  

Objective Evaluation of Speech Timbre Based on Auto Encoder

作　　者：涂中文[1] 赵艳明[2] 宋金宝[2] TU Zhong-wen;ZHAO Yan-ming;SONG Jin-bao(School of Presentation Arts,Communication University of China,Beijing 100024,China;School of Information Engineering,Communication University of China,Beijing 100024,China)

机构地区：[1]中国传媒大学播音主持艺术学院,北京100024 [2]中国传媒大学信息工程学院,北京100024

出　　处：《中国传媒大学学报（自然科学版）》2017年第4期8-13,共6页Journal of Communication University of China：Science and Technology

摘　　要：本文详细介绍了嗓音识别和深度学习的基本原理,然后阐述了怎样将深度学习理论应用于嗓音质量识别分析的研究中,从基于深度学习的语音特征参数提取和神经网络模型建模两方面入手,首先提取不同的音频特征参数,然后构建以稀疏编码器为核心的堆栈式自动编码器,"封顶"softmax分类器以构成完整的深度学习网络,最后测试了不同特征参数、不同的网络层数和网络节点数对于实验准确率的影响。This paper would first introduce the basic principles of voice recognition and Deep Learning in detail,then set forth how to put the theory of DL into use of voice quality recognition.Starting with extraction of the feature of audio signal,we then set up stacked auto encoder with sparse coding as the core and softmax classification as top.Finally,we take the different features as input and change the number of hidden layers and hidden unites to observe their impacts on validating accuracy.

关 键 词：语音音色 客观评价 自动编码器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]