摘要：

利用小样本声纹作为训练集训练卷积神经网络(CNN)时,网络不能达到较好的收敛状态,从而导致识别率较低。为此,提出一种新的声纹识别方法。利用深度CNN提取潜在的声纹特征,在CNN训练过程中采用基于凸透镜成像原理的图像增多算法解决小样本训练样本不足的问题,并在卷积过程中引入快速批量归一化(FBN)方法以提高网络收敛速度、缩短训练时间。在包含630人的TIMIT语音数据库中进行训练、验证和测试,结果表明,FBN-Alexnet网络比Alexnet网络训练时间缩短48.2%,与GMM、GMM-UBM及GMM-SVM方法相比,该方法识别率分别提高7.3%、2.2%、2.8%。

关键词: 声纹识别, 深度学习, FBN-Alexnet网络, 小样本, 快速批量归一化, 图像增多算法

Abstract:

When training Convolutional Neural NetWork(CNN) with small sample voiceprints as training set,the network cannot reach a good convergence state,which results in low recognition rate.So,this paper proposes a new voiceprint recognition method.The proposed method uses deep CNN to extract the rich and latent features of voiceprint,which improves the voiceprint recognition rate.In order to solve the problem that small sample cannot train the CNN,this paper proposes an image increasing algorithm based on the principle of convex lens imaging.At the same time,the Fast Batch Normalization (FBN) is introduced in the convolutional process,which improves the speed of the network convergence and shortens the training time.Select a TIMIT speech database containing voices of 630 speakers for training,validating and testing.Experimental results show that,compared with the GMM,GMM-UBM,and GMM-SVM algorithms,the proposed method improves the recognition rate by 7.3%,2.2%,and 2.8% and compared with the original network,the training time of the FBN-Alexnet network is reduced by 48.2%.It means that it is an effective method for voiceprint recognition of small samples.

Key words: voiceprint recognition, deep learning, FBN-Alexnet network, small sample, Fast Batch Normalization (FBN), image increasing algorithm

中图分类号: 

• TN912.34