Research On Audio Recognition Method Based On Residual Network And Random Forest

Environmental Sound Classification(ESC)is one of the important branches in the field of audio processing.ESC tasks have important application scenarios in robot navigation,audio retrieval,audio forensics,and other context-aware and wearable devices.The difficulty of the environmental sound recognition task is mainly that the environmental sound information is often mixed with a large amount of useless and random sound information,which makes the sound features with very similar similarity often appear in different sound scenes.There is still much room for improvement in the classification accuracy of existing methods in ESC tasks.In this thesis,an audio recognition algorithm based on residual network and random forest is proposed.The deep learning model is used to extract the features of the audio.By transforming the one-dimensional time domain audio signal into a two-dimensional frequency domain signal,the residual network model is utilized to extract feature,which is performed to improve the quality of audio feature extraction.For the audio dataset,because the recording of the sound event is very difficult,the data volume is small,and the training deep learning model is easy to over-fitting,which leads to the degradation of precision.This thesis designs a method based on the combination of residual network and random forest.our method has effective extraction of audio features and alleviates the phenomenon of over-fitting,so this method not only improves the accuracy of audio recognition,but also accelerates the efficiency of prediction.In this thesis,a representative environmental sound data set is used,and the proposed audio recognition method is used for training prediction on these data sets.The experimental results show that the recognition accuracy of the proposed method is greatly improved,and the vocal data set of the actual recording is also well verified,which verifies the universality and effectiveness of the method.