基于强约束字典联合深度神经网络的单通道语音分离

Single⁃channel speech separation based on strongly constrained dictionary combined with a deep neural network

针对基于字典学习语音分离方法的“交叉投影”问题,提出了强约束的优化函数,不仅抑制重构信号和目标信号的误差,约束干净信号在联合字典上的误差,而且抑制干净信号在其他字典上的投影并限制字典间的原子相关性。此外,为了进一步提高两个相似信号的分离效果,提出基于强约束字典联合深度神经网络的单通道语音分离方法,首先利用强约束字典实现目标与干扰语音的初步分离,然后通过联合约束利用深度神经网络实现语音与干扰语音交叉投影残余的分离。实验结果表明,与其他优秀单通道语音分离方法相比,该算法有效提升了语音分离系统的性能。

Aiming at the‘cross projection’problem of speech separation methods based on dictionary learning,a strongly constrained optimization scheme is proposed.This scheme suppresses the error of the reconstructed signal and the target signal,constrains the error of the clean signal on the joint dictionary,and suppresses the clean signal in other projections on dictionaries as well as atomic dependencies between dictionaries.Further,in order to improve the separation effect of two similar signals,a single⁃channel speech separation method based on a strongly constrained dictionary combined with a deep neural network is proposed.The strong constraint dictionary is used to realize the preliminary separation of target speech and interference speech,and then the deep neural network is used through joint constraints.The network achieves the separation of the cross⁃projection residuals of the speech and the interfering speech.Experiments show that compared with other excellent single⁃channel speech separation methods,