基于BP神经网络的耳语音转换为正常语音的研究  

Reconstruction of Normal Speech from Whispered Speech based on BP Neural Network

作　　者：韩韬[1] 陶智[1] 顾济华[1] 赵鹤鸣[2] 李玲[1] 

机构地区：[1]苏州大学物理科学与技术学院,江苏苏州215006 [2]苏州大学电子信息学院,江苏苏州215021

出　　处：《通信技术》2009年第2期152-155,共4页Communications Technology

基　　金：国家自然科学基金(the National Natural Science Foundation of China under Grant No.60572076);江苏省高校自然科学研究计划基金资助项目(No.05JKB510113);国家级大学生创新性实验计划项目(No.57315755)。

摘　　要：提出了一种基于BP神经网络的汉语耳语音转换为正常语音的方法。首先提取正常语音、耳语音的共振峰参数,使用BP神经网络训练出耳语音到正常语音共振峰参数的转换模型;然后根据模型求出与耳语音对应的正常语音共振峰参数,采用共振峰合成的方法将耳语音转换为正常语音。实验结果表明:使用该方法转换的正常语音DRT得分为80%,MOS得分为3.5,在可懂度和音质方面均达到了满意的效果。This paper tells of a new approach for reconstructing normal speech from Chinese whispered speech based on BP neural network. First, the formants of normal speech and whispered speech are acquired, and the BP neural network is used to get a model of the conversion from the whispered speech to the formants of normal speech, then the formant of whispered speech is converted by adopting this model. Finally, the whispered speech is converted into the normal speech by using the formant synthesizer. Simulation results show that the score of the DRT of the converted speech is 80% and the MOS is 3.5, both intelligibility and quality of the converted speech are satisfied.

关 键 词：耳语音 语音转换 BP神经网络 共振峰 

分 类 号：TN912.33[电子电信—通信与信息系统]