Research On Recognition Of Sound Events Based On Multi-scale And Multi-level Feature Analysis

Sound is one of the most important ways to transmit information.It has become an important research topic to detect and identify the sound events contained in audio automatically by machine,which plays a key role in environmental monitoring,unmanned driving,ocean monitoring,robot,health care and other fields.The general process of sound event recognition includes feature extraction and feature selection,classification algorithm design,performance measurement and model evaluation.With the vigorous development of large data and artificial intelligence technology,data-driven with deep learning method has become a mainstream method of sound event recognition.Because of the diversity of sound event categories,different kinds of sound events often have different event scales and feature levels,which challenges the research of sound event recognition methods.Based on the analysis of the general methods of sound event recognition,a multi-scale and multi-level feature analysis method is designed and put forward in this paper.Taking the air and underwater sound data as the research objects,the data pretreatment,the design and selection of deep learning model,the training of model and the analysis of experimental results are analyzed.Aiming at the diversity of the feature and time scale,a multi-scale and multi-level feature analysis method based on convolutional recurrent neural network is proposed in this paper,which considers the cascade of multi-scale features from the feature input and adopts the cascade of multi-level features between network levels.Hierarchical feature connection can enhance the effectiveness of feature extraction,improve the detection and recognition accuracy,slow down the over-fitting of the model and enhance the stability of the model.The experimental results based on air sound and underwater sound datasets show that the sound event recognition model using multi-scale multi-level feature analysis has a better recognition accuracy than the single-scale hierarchical feature model.