深度迁移模型下的小样本声纹识别方法

Voiceprint recognition method of small sample based on deep migration model

基于神经网络的生物特征识别通常需要大量训练样本,目标数据集不足使得神经网络难以获得实际应用。在小样本声纹识别中,为提高识别的准确率,结合卷积神经网络与迁移学习提出一种基于小样本的说话人识别方法。在卷积过程中引入快速批量归一化(fast batch normalization,FBN),提高深度神经网络收敛速度。将预训练模型中的全连接层改为RBM(restricted Boltzmann machine),用小样本声纹训练RBM和分类器,RBM能够进一步学习小样本声纹特有的高阶特征,消除在迁移过程中声纹数据集间的差异。选取包含400人的AISHELL-ASR0009-OS1语音数据库及实验室自采的20人语音库进行实验,实验结果表明,融合FBN的神经网络相比原始网络的训练时间减少了35.6%,最优方法相比其它两种方法识别率提高了9.7%-41.3%,验证了所提方法的可行性和有效性。

Biometric identification based on neural network usually requires a large number of training samples,and the lack of target data sets makes it difficult for neural networks to obtain practical applications.Combining convolutional neural network and transfer learning,a method of small sample speaker recognition was proposed to improve the recognition accuracy.FBN was introduced to improve the convergence speed of deep neural networks.The full connected layer was modified by RBM(restricted Boltzmann machine),RBM and classifier were trained with small sample voiceprint.RBM could further study the high-order features unique to small sample voices,and the difference between the vocal data sets during the transfer process was eliminated.AISHELL-ASR0009-OS1voice database containing voices of400speakers and the self-collected voice dataset containing voices of20speakers were selected to perform the experiment.The results show that the normalized network is35.6%shorter than that of the original network.The optimal method improves the recognition rate by9.7%-41.3%compared with the other two methods,verifying the feasibility and effectiveness of the proposed method.