摘要
为了降低双耳语音中噪声和混响的影响,提高语音质量和可懂度,提出了一种基于注意力机制和改进的卷积循环神经网络的双耳语音增强算法.在该算法中,首先提取双耳语音的谱特征和双耳线索,对谱特征应用通道注意力得到可靠的谱特征,同时对双耳线索应用空间注意力得到可靠的双耳线索作为神经网络的输入特征.然后,构建了将模型注意力作为卷积循环神经网络编解码层的跳跃连接的神经网络结构,并利用双向长短期记忆网络获取时序信息.实验结果表明:在不同噪声与混响的条件下,所提出的算法具有更好的性能.
In order to reduce the influence of noise and reverberation in binaural speech,and improve speech quality and intelligibility,a binaural speech enhancement algorithm based on attention mechanism and improved convolutional recurrent neural network was proposed.In this algorithm,the spectral features and binaural cues of binaural speech were first extracted,and channel attention was applied to the spectral features to obtain reliable spectral features,while reliable binaural cues were obtained by applying spatial attention to the binaural cues,then the two features were combined as the input of neural network.A neural network structure that uses model attention as a skip connection for the encode layer and decode layer of convolutional recurrent neural network,and the bidirectional long and short-term memory network was used to obtain time domain information.Experimental results show that the proposed algorithm has better performance in different noise and reverberation conditions.
作者
李如玮
李秋艳
赵丰年
刘尚枫
LI Ruwei;LI Qiuyan;ZHAO Fengnian;LIU Shangfeng(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2023年第9期125-131,166,共8页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金项目资助(61971016).
