基于自动编码生成对抗网络的语音增强算法

Speech enhancement based on auto-encoders and generative adversarial network

针对当前基于深度学习的语音增强算法中忽略语音相位作用的问题,提出一种基于自动编码生成对抗网络的语音增强算法。采用一种由自动编码器(AE)与生成对抗网络(GAN)相结合的综合学习框架,在语音波形层面进行操作即端到端处理,充分利用时域上的信息。AE自动提取语音特征,有监督的学习带噪语音与纯净语音之间的非线性关系,将语音建模为概率模型中标签和潜在属性的组合;在反向传播时,判别网络和分类器采用交叉熵损失函数,生成网络采用平均差异损失函数,这种不对称损失函数使GAN训练更加稳定。增强后的样本验证了所提算法的可行性,客观评估验证了其有效性,整体性能优于DNN的算法。

Aiming at the problem of neglecting the phase effects of speech in the current deep learning-based speech enhancement algorithm,a speech enhancement algorithm based on auto-encoder and conditional generative adversarial network(AE-CGAN)was proposed.A comprehensive learning framework combining auto-encoder(AE)and generative adversarial network(GAN)was adopted,and it operated at the speech waveform level,that was,end-to-end processing,making full use of information in the time domain.AE automatically extracted speech features,supervised learning the nonlinear relationship between noisy and clean speech,and modeled speech as a combination of labels and latent attributes in a probabilistic model.In the back propa-gation,the discriminator network and the classifier used the cross entropy loss function,and the generator network used the average difference loss function.This kind of asymmetric loss function made the GAN training more stable.The test set was used to evaluate the trained model.The enhanced sample confirmed the feasibility of the proposed algorithm,and the objective evaluation confirmed its validity.