基于深度神经网络的说话人年龄分类研究

Speaker age classification based on deep neural network

说话人年龄分类是通过说话人的语音来估测说话人的年龄范围,属于说话人属性分析的重要内容。传统说话人年龄分类是利用人工提取特征加后端分类器的方法,这种方法不一定能够得到最优的分类效果。针对该问题,利用深度神经网络自动获取特征表示的能力,提出一种基于深度神经网络的年龄分类方法,采用提取文本无关的说话人语音特征来通过深度神经网络对说话人年龄进行分类。在1000 h的AISHELL⁃2中文开源语音数据库上验证了所提出方法的有效性。与基于i⁃vector结合后端分类器的两阶段说话人年龄分类方法做了比较,实验结果证明,与i⁃vector方法相比,识别正确率获得了3.08%的性能提升。

The speaker age classification refers to the estimation of his(or her)age range according to the speaker's speech,which is an integral part of the speaker attribute analysis.The traditional method of age classification for speakers is to use manual feature extraction and back⁃end classifier.However,this method may be unable to obtain the optimal classification result.To solve this problem,an age classification method based on deep neural network is proposed by utilizing the ability of the deep neural network to automatically obtain feature representations.The features independent from any texts is extracted with the method to classify the speaker′s age by the deep neural network.The effectiveness of the proposed model was verified on the 1000 hour AISHELL⁃2 Chinese open⁃source speech database.The experiments result show that,in comparison with the two⁃stage speaker age classification method based on I⁃vector and back⁃end classifier,the recognition correctness of the propose model has increased 3.08%.