摘要
使用深度学习技术进行语音分离已经取得了优异的成果。当前主流的语音分离模型主要基于注意力模块或卷积神经网络,它们通过许多中间状态传递信息,难以对较长的语音序列建模导致分离性能不佳。首先提出了一种端到端的双路径语音分离网络(DPCFNet),该网络通过引入改进的密集连接块,使编码器能提取到丰富的语音特征。然后使用卷积增强Transformer(Conformer)作为分离层的主要组成部分,使语音序列中的元素可以直接交互,不再通过中间状态传递信息。最后将Conformer与双路径结构相结合使得该模型能够有效地进行长语音序列建模。实验结果表明,相比于当前主流的Conv-Tasnet算法及DPTNet算法,所提出的模型在信噪失真比(Signal to noise Distortion Ratio,SDR)和尺度不变信噪失真比(Scale-Invariant Signal to noise Distortion Ratio,SI-SDR)上有明显提高,分离性能更好。
The deep learning techniques used in speech separation have yielded excellent results.The current mainstream speech separation models are mainly based on attention modules or convolutional neural networks,which transmit information through many intermediate states,which is difficult to model longer speech sequences and can lead to poor separation performance.First,this paper proposes an end�to-end dual-path speech separation network(DPCFNet),which enables the encoder to extract rich speech features by introducing an improved densely connected block.Then,it uses a Conformer(convolution�augmented transformer)as the main part of the separation layer so that elements in the speech sequence can interact directly instead of passing information through intermediate states.Finally,Conformer and dual�path structure are combined to make the model effective for modeling long speech sequences.Experimental results indicate that compared with the current mainstream Conv-Tasnet algorithm and DPTNet algorithm,the proposed model has obvious improvement in SDR(Signal to noise Distortion Ratio)and SI-SDR(Scale�Invariant Signal to noise Distortion Ratio),and has better separation performance.
作者
王钧谕
高勇
WANG Junyu;GAO Yong(Sichuan University,Chengdu Sichuan 610065,China)
出处
《通信技术》
2023年第5期585-589,共5页
Communications Technology
