基于特征融合的正常语音和低语语音分类系统  

Classification system of normal and whisper speech based on feature fusion

作　　者：王睿[1] 艾斯卡尔·艾木都拉[2] WANG Rui;ASKAR Hamdulla(School of Information Science and Engineering,Xinjiang University,Urumqi 830017,China;Key Laboratory of Multilingual Information Technology,Xinjiang University,Urumqi 830017,China)

机构地区：[1]新疆大学信息科学与工程学院,新疆乌鲁木齐830017 [2]新疆大学多语种信息处理实验室,新疆乌鲁木齐830017

出　　处：《现代电子技术》2023年第4期161-165,共5页Modern Electronics Technique

基　　金：国家重点研发计划项目(2017YFC0820602)。

摘　　要：低语(Whisper Speech)是指凑近别人耳朵小声说话,即私下里轻微的说话声。低语是一种常见的发音方式,由于发音方式比较特殊,其与正常语音在特征方面有较大差异。目前,区分低语和正常语音多数是借助于各类软件对某些声学特征进行直接观测,比如常见的频谱图等,而对于二者的分类模型研究较少,且没有一个公开的汉语低语语料库。为此,首先创建一个汉语低语语料库;其次,建立一种鲁棒的低语与正常语音的分类系统,提出一种基于卷积神经网络(CNN)的特征融合方法,该方法将光谱平坦度(Spectral Flatness)和语音均方根(Root Mean Square,RMS)相结合。实验结果表明,所提出的特征融合方法能够提高低语与正常语音分类系统的性能,与基线模型相比,准确率提高21.67%。Whisper Speech refers to whispering close to others′ ears,that is,speaking quietly in private. Whispering is a common way of pronunciation. Due to the special way of pronunciation,it is quite different from the normal voice in terms of features. The distinction between whispers and normal speech is mostly based on the direct observation of some acoustic features by means of various software,such as common spectrograms,while there is little research on the classification model of the two,and there is no public Chinese whispers corpus. The corpus of Chinese whispers is created and then a robust classification system of low speech and normal speech is established. A feature fusion method based on CNN(convolutional neural network)is proposed,which combines spectral flatness and RMS(root mean square)of speech. The experimental results show that the proposed feature fusion method can improve the performance of the whisper and normal speech classification system,and its accuracy is increased by 21.67% compared with the baseline model.

关 键 词：低语 声学特征 汉语语料库 分类器 卷积神经网络 光谱平坦度 均方根 特征融合 

分 类 号：TN912-34[电子电信—通信与信息系统]