Abstract: At present,most of the research on sound recognition and detection is based on the datasets with strong labels.Howe-ver,in real-world sound recognition and detection tasks,it is difficult to obtain strong label audio data due to incomplete audio labels with a large amount of noise,which in turn affects the accuracy of sound identification and detection.To this end,a multi-scale attention fusion mechanism is proposed based on the convolutional cyclic neural network model.This mechanism uses the attention gating unit to make more use of the effective features while reducing the effects of noise in the sound time-frequency map features.At the same time,feature fusion is performed by combining convolution kernels of multiple sizes to further improve the effective extraction of sound features.In addition,the sound signal is identified by a weighting method that combines the results of the frame detection.Finally,using the proposed model,a weak labeled data set containing 17 kinds of urban vehicle sounds is selected from the AudioSet database for detection and identification in the weak label environment.For the test set,the F1 value of sound recognition result is 58.9%,and the F1 value of detection result is 43.7%.The simulation experiments show that the CRNN baseline model used in this paper is more accurate than the traditional sound recognition detection model under the weakly labeled city vehicle sound datasets.And the methods involved in the paper,such as the importance weighted recognition method and multi-scale attention fusion method,can improve the accuracy of the model for sound recognition and detection.

Key words: Attention, Multi-scale, Sound detection, Sound recognition, Weak label