基于生成对抗网络的语音信号分离

Speech Signal Separation Based on Generative Adversarial Networks

基于深度学习的单声道语音分离需要计算时频掩蔽,但现有语音分离方法中时频掩蔽不可学习,也未将其封装到深度学习中进行优化,通常依赖于维纳滤波法进行后续处理。为此,提出一种基于生成对抗网络的语音信号分离方法。在语音生成阶段引入递归推导算法和稀疏编码器来改进时频掩蔽生成结果,并将生成的语音输入至判别器进行分类,以降低信号源之间的扰动。实验结果表明,与基于深度神经网络的语音信号分离方法相比,该方法的SDR、SIR分离指标分别提高6.2 dB和5.0 dB。

The single-channel speech separation based on deep learning needs to calculate the time-frequency masking,which,however,cannot be learnt in the existing methods.Moreover,the time-frequency masking is not encapsulated in in-depth learning for optimization,so it relies on Wiener filtering for subsequent processing.Therefore,this paper proposes a speech signal separation method based on Generative Adversarial Networks(GAN).In the speech generation stage,the recursive derivation algorithm and sparse encoder are introduced to improve the time-frequency generation results.Then,the generated speach is eatered into the discriminator for classification,so as to reduce the disturbance between signal sources.The experimental results show that compared with other speech signal separation methods,such as the codec-based method and the recurrent neural network-based method,the SDR and SIR separation indexes of the proposed method increase by 6.2 dB and 5.0 dB respectively.