基于U-Conformer的多特征融合鸟鸣声分离方法

U-Conformer-based multi-feature fusion bird sound separation

针对多个鸟类个体同时发声导致的鸣声混叠问题,本文提出了一种融合录音通道间空间特征的鸟类鸣声分离方法.该方法将混叠鸣声信号的声谱特征和空间特征作为分离模型的输入,提出深度学习模型U-Conformer来预测每个鸣声源方向的幅值谱掩膜(spectral magnitude mask,SMM),通过模型估计的SMM从混叠鸣声信号中恢复每个鸣声源信号.由多源混叠鸟类鸣声数据的实验结果表明,本文提出的分离方法较其他深度学习模型结构具有更好的分离效果,有助于更好地分析野外鸟类鸣声录音.

Simultaneous vocalization of multiple birds leads to overlapping bird sound.In this paper a bird sound separation method,with integrated spatial features,is proposed.In this method,both spectral and spatial features of overlapped sound signals are used as input,U-Conformer is used as a separation model to predict spectral magnitude mask(SMM).The sound source signal is recovered from mixed sound signal by estimated SMM.The generated multi-channel bird sound data confirm that this method has better performance in bird sound separation compared with existing methods.