Aiming at the fact that the existing sound event detection methods do not pay enough attention to different time and frequency band information,and the traditional single feature cannot characterize the spatial phase information of time-frequency overlapping sound events,a time-frequency attention algorithm based on dual feature input was proposed,in which the logarith-mic Mel spectrum and the generalized cross-correlation with phase transformation were used as input,and the attention mecha-nism was used to capture more effective time-frequency features from the two dimensions of time and frequency,respectively.To improve the multi-resolution processing ability of the algorithm,the feature pyramid model based on attention was designed to learn multi-scale features and to help the model identify different sound events.Experimental results show that the proposed algorithm can effectively extract key features and perform multi resolution processing,improving the performance of sound event detection.
维普
万方数据