基于神经网络噪声分类的语音增强算法  

Speech Enhancement Algorithm Based on Neural Network Noise Classification

作　　者：张行 赵馨[1] ZHANG Hang;ZHAO Xin(ChangChun University of Science and Technology,ChangChun 130022,China)

机构地区：[1]长春理工大学,吉林长春130022

出　　处：《中国电子科学研究院学报》2020年第9期880-885,893,共7页Journal of China Academy of Electronics and Information Technology

基　　金：吉林省发改委项目(2018c035-3)。

摘　　要：传统的语音增强算法由于缺少背景噪声信息,在进行语音处理时对不同的语音信号采取相同的处理方式,因此存在估计噪声不准确、增强语音失真及噪声抑制不明显等问题,最终导致语音的增强效果不明显。在此基础上提出一种根据不同噪声进行参数自适应的语音增强算法,首先,通过神经网络进行精确分类;然后,根据分类结果选取不同参数的IMCRA算法进行噪声估计;最后,采取OMLSA算法对语音信号进行增强。实验结果表明,经噪声分类后的增强算法能够取得更好的增强效果,更多的保留语音信号中的信息,且能够在不降低语音可懂度的同时提高语音的质量。Due to the lack of background noise information in traditional speech enhancement algorithms,different speech signals are processed in the same manner during speech processing.Therefore,there are problems such as inaccurate noise estimation,enhanced speech distortion and insignificant noise suppression,which eventually lead to insignificant speech enhancement effect.On this basis,propose a parametric adaptive speech enhancement algorithm based on different noise,Firstly,the noise is accurately classified by the neural network,and the noise is estimated by IMCRA algorithm with different parameters selected according to the noise type,Finally,OMLSA algorithm is adopted to enhance the speech signal.The experimental results show that the enhanced algorithm after noise classification can achieve better enhancement effect,retain more information in speech signal,and improve speech quality without decreasing speech intelligibility.

关 键 词：神经网络 音频增强 噪声分类 IMCRA算法 OMLSA算法 

分 类 号：TN912.35[电子电信—通信与信息系统]