面向中文语音情感识别的改进栈式自编码结构

发布时间：2018-01-14 20:26

本文关键词：面向中文语音情感识别的改进栈式自编码结构　出处：《东南大学学报(自然科学版)》2017年04期 　论文类型：期刊论文

【摘要】：为进一步提高汉语语音情感识别率,基于深度学习中的自编码、降噪自编码及稀疏自编码的网络结构,提出了一种改进的栈式自编码结构.该结构第1层使用降噪自编码学习一个比输入特征维数更大的隐藏特征,第2层采用稀疏自编码学习稀疏性特征,最后使用softmax分类器进行分类识别.训练过程首先采用逐层预训练的方法,达到网络参数全面初始化的目的,然后对整个网络进行微调.在中文语音库上的情感识别实验显示,相较于单独使用栈式降噪或稀疏自编码,所提结构具有更好的识别效果.此外,基于CASIA库的对比实验显示,该结构比K近邻算法、稀疏表示方法、传统支持向量机和人工神经网络识别率分别提高了53.7%,29.8%,14.3%和1.9%.在自行录制的语音库中,该结构的识别率比人工神经网络提高了1.64%.
[Abstract]:In order to improve the emotion recognition rate of Chinese speech, the network structure is based on self-coding, de-noising self-coding and sparse self-coding in depth learning. An improved stack self-coding structure is proposed, in which layer 1 uses noise reduction self-coding to learn a hidden feature with larger dimension than input feature, and layer 2 uses sparse self-coding to learn sparse feature. Finally, the softmax classifier is used for classification recognition. Firstly, the method of pre-training layer by layer is used in the training process to achieve the goal of the overall initialization of network parameters. Then the whole network is fine-tuned. The experiment of emotion recognition on Chinese speech corpus shows that the proposed structure has better recognition effect than using stack noise reduction or sparse self-coding alone. The contrast experiments based on CASIA library show that the recognition rate of this structure is increased by 53.7% than that of K-nearest neighbor algorithm, sparse representation method, traditional support vector machine and artificial neural network, respectively. In the self-recorded corpus, the recognition rate of the structure is 1.64% higher than that of artificial neural network.
【作者单位】：东南大学水声信号处理教育部重点实验室;南京工程学院通信工程学院;
【基金】：国家自然科学基金资助项目(61375028,61571106,61673108) 江苏省青蓝工程资助项目江苏省博士后科研资助计划资助项目(1601011B) 江苏省“六大人才高峰”资助项目(2016-DZXX-023) 中国博士后科学基金资助项目(2016M601695)
【分类号】：TN912.34
【正文快照】： 语音是人机交互最自然友好的方式之一,承载着说话人丰富的情感信息.但如今的人机交互并不能像人与人交流那样自由,其原因是机器不能像人一样判断对方话语中的情感.语音情感识别在人工智能领域有着非常广阔的未来,其终极目标是让机器能够像人类一样通过语音识别人类情感,实现更

【参考文献】