多尺度卷积的时频域语音分离方法研究

Speech separation in time-and-frequency domain based on multi-scale convolution

在进行混合语音分离时,信号时域特征的深度学习语音分离性能优于频域特征。但目前时域特征的语音分离方法在真实噪声环境下的鲁棒性较差,且单一时域特征对分离模型的性能存在局限性。因此,提出一种基于Conv-TasNet网络的多特征语音分离方法,融合频域特征与时域特征,提高数据的多维信息。为了进一步提高分离网络性能,引入多尺度卷积块,提高网络对特征的提取能力。在包含真实噪声的实验环境下,所提方法与Conv-TasNet模型和最新的时频域融合语音分离基线模型相比,性能分别提高了0.91和0.52 dB,有效提升了语音分离的性能及鲁棒性。

In mixed speech separation, the performance of signal time-domain features is better than that of frequency-domain features. However, the current speech separation methods based on time domain feature have poor robustness in real noise environment, and single time domain feature has limitations on the performance of the separation model. Therefore, a multi-feature speech separation method based on Conv-TasNet network is proposed, which integrates frequency domain features and time domain features to improve multidimensional information of data. In order to further improve the performance of separation network, multi-scale convolution block is introduced to improve the feature extraction ability of network. Compared with the Conv-TasNet model and the latest time-frequency fusion speech separation baseline model, the performance and robustness of the proposed method are improved by 0.91 and 0.52 dB respectively in the experimental environment containing real noise.