基于噪声特性的语音增强算法

Speech Enhancement Algorithm Based on the Noise Characteristics

针对不同的语音增强算法对不同噪声的增强效果不同,提出了一种基于深度神经网络的噪声分类的语音增强算法。首先,使用深度神经网络(DNN)算法对噪声进行分类。分类算法包括训练阶段和分类阶段。在训练阶段,采用babble,car,street,train四中噪声对DNN进行训练;在分类阶段,将提取的噪声输入训练好的DNN中,得到分类结果,并对分类性能进行评估。其次,采用PESQ,LSD及SNR等语音评估方法,对不同的含噪语音在不同信噪比、不同语音增强算法下进行评估。语音增强算法包括子空间法、维纳滤波算法、谱减法及对数最小均方误差法(log MMSE),噪声包括babble,car,street,train,信噪比为-5db,0db和5db,并对通过评估得到的值采用平均值法得到噪声和语音增强算法的最佳匹配;最后,针对不同分类噪声,采用不同的增强算法进行语音增强,并对4种噪声之外的噪声根据本文算法选取相应的语音增强算法。

According to that different speech enhancement algorithm has a different enhancement effect on different noise type,a kind of speech enhancement algorithm based on the noise classification of deep neural network was proposed. Firstly,deep neural network（ DNN） algorithm was exploited to classify noise and the classification algorithms comprised two stages： the training stage and classification stage. In the training stage,the noise babble,car,street and train were using for training DNN,in classification stage,inputting the extracted noise to the trained DNN,obtained the classification result and evaluated the classification performance; Secondly,speech evaluation methods such as PESQ,LSD and SNR were adopted to evaluate the performance. For different noisy speech in different SNR,different speech enhancement algorithms were evaluated. The speech enhancement algorithms which were adopted including subspace,wiener filtering,spectral subtraction and log MMSE algorithm,the adopted noise including babble,car,street and train,the adopted SNR including- 5 db,0 db and 5 db. The values which were obtained by evaluated were used to get the best match between noise and speech enhancement algorithm; Finally,according to different noise type,using different speech enhancement algorithm to conduct speech enhancement,as for another noise type,according to proposed algorithm select relevant speech enhancement algorithm.