深度复卷积递归网络模型的师生学习语声增强方法

Teacher-student learning for speech enhancement based on deep complex convolution recurrent network

基于深度神经网络的方法已经在语声增强领域得到了广泛的应用,然而若想取得理想的性能,一般需要规模较大且复杂度较高的模型。因此,在计算资源有限的设备或对延时要求高的环境下容易出现部署困难的问题。为了解决此问题,提出了一种基于深度复卷积递归网络的师生学习语声增强方法。在师生深度复卷积递归网络模型结构中间的复长短时记忆递归模块提取实部和虚部特征流,并分别计算帧级师生距离损失以进行知识转移。同时使用多分辨率频谱损失以进一步提升低复杂度学生模型的性能。实验在公开数据集Voice Bank Demand和DNS Challenge上进行,结果显示所提方法相对于基线学生模型在各项指标上均有明显提升。

Deep learning-based methods have been widely used in the field of speech enhancement.However,a model with large scale and high complexity is typically required to achieve the desired performance.Hence,deployment difficulties may occur in devices with limited hardware resources or in applications with strict latency requirements.In order to solve this problem,a teacher-student learning method for speech enhancement based on deep complex convolution recurrent network(DCCRN)is proposed.The real and imaginary feature streams are extracted from the output of the complex long short term memory(Complex LSTM)in the middle of the DCCRN model,and the frame-level teacher-student distance loss is calculated to transfer knowledge.Meanwhile,the multi-resolution short-time Fourier transform loss is used to further improve the performance of the low-complexity student model.The experiment was conducted on the open source dataset Voice Bank Demand and DNS Challenge,and the results show that the proposed method has a significant improvement in various indicators compared with the baseline student model.