Research On Feature Extraction And Recognition Of Sound Event

Sound waves caused by physical event can be defined as sound event.The feature extraction and recognition of sound event can help us obtain environmental information to guide production and life.The process of sound event recognition is divided into two parts: feature extraction and recognition.At present,the mainstream feature extraction methods are mainly based on traditional speech features,and lack pertinence for sound event.In addition,due to environmental factors,it is difficult for traditional algorithms to mining features from complex sound event.The neural network performs better,but requires sufficient data.Based on the research of sound feature extraction and classifier,this thesis proposed a new method for sound event recognition.This thesis selected 5 major categories(50 minor categories)of non-speech sound events as research data set,including animals,natural soundscapes and water sounds,human non-speech sounds,interior/domestic sounds,exterior/urban noises.Based on Harmonic and Percussive Source Separation,multiple features including Log Mel Spectrogram,Mel Frequency Cepstral Coefficients,Chromagram,Zero Crossing Rate and their first and second order differences were extracted and combined.A residual neural network was designed for sound event classification,as well as audio sample amplification and data enhancement method based on mix up and random crop.Also,the hyperparameter tuning of SGD optimizer was analyzed.The data set was divided into training set and validation set according to a proportion of 80% and 20%.After 5-fold cross validation,an average classification accuracy of 88.3% was reached,which exceeds the manual classification accuracy of 81.3%,indicating that the method proposed in this thesis has a good performance for sound event recognition.