Deep Learning Based Sound Recognition Classification System

Sound is everywhere.After hearing the sound,people continuously process and understand the audio consciously or subconsciously,so as to provide us with information about the surrounding environment.Intelligent environmental sound classification is a research field that is constantly evolving in many practical applications.Although there has been a lot of research in the audio field(such as speech and music),relatively little work has been done to classify sounds in the environment.The use of deep learning to classify sound image processing has not yet appeared,which leads to the use of convolutional neural networks to classify discrete sound signals that occur over time.Intelligent environmental sound classification is a research field that is constantly evolving in many practical applications.Although there has been a lot of research in the audio field(such as speech and music),relatively little work has been done to classify sounds in the environment.The use of deep learning to classify sound image processing has not yet appeared,which leads to the use of convolutional neural networks to classify discrete sound signals that occur over time.This research is to apply deep learning technology to the classification of sounds in the environment,especially focusing on the recognition of sounds in life,that is,using deep learning technology to classify the sounds in our lives.When given an audio sample in a computer-readable format(such as a.wav file)that lasts a few seconds,it is desirable to be able to determine whether it is one of the sounds in the data set and the corresponding likelihood score.Conversely,if the target sound is not detected,we will get an unknown score.The main research contents are as follows:(1)The meaning of sound classification and its application in realityAs a carrier of information,sound is an indispensable element in human society.It is included in all aspects of human life.Although the human ear can effectively recognize some sounds,its ability is limited in more complex situations,so an intelligent sound classification system is needed to assist humans in recognizing sounds in order to achieve certain goals.At present,there are many application requirements for voice recognition classification,such as assisting deaf-mute people in their daily activities,cars that can recognize sounds inside and outside,and predictive maintenance of machines.These applications will help improve people’s lives and increase people’s work efficiency.(2)Research the convolutional neural network(CNN)in deep learning technology to perform high-precision and large-scale classification of images after sound image processingThis topic uses CNN technology in deep learning to classify sounds in the living environment,image the collected sound data set according to needs,and then preprocess each frame of sound data of the sound image,and use the Mel spectrum cepstrum(MFCC)to extract the sound features needed for training the model,segment these data sets with feature labels,and store the classification labels together in the Dataframe in Panda,and then put the data set into the established.Training in the sound classification model.(3)Optimized the algorithm for sound classification accuracyThe establishment of the sound classification model involves the application of deep learning algorithms,and the general benchmark algorithm cannot meet the corresponding accuracy requirements,so it is necessary to improve the algorithm in the corresponding model to meet the requirements of improving the recognition accuracy.This article is in five types Based on the benchmark model algorithm(decision tree,KNN,random forest,support vector machine,and majority voting algorithm),through the improvement of MLP(multilayer perceptron)model and CNN(convolutional neural network)model,the hierarchical structure,The feature-related parameters and the number of nodes in each layer of the model are modified to create a voice recognition system with higher classification accuracy.(4)System test and algorithm comparison analysis researchInput 8733 wav files with a duration of a few seconds into the optimized voice recognition classification system.The system can accurately match the input voice with the voice learned in the model.Each voice will be given a score with the highest score.The sound category is the target sound category.At the same time,the sample sound is continuously trained with the above five algorithms and the improved MLP and CNN models.Among the five benchmark algorithms,the highest recognition accuracy is 68% for SVM,and the recognition accuracy reaches 88% and 92% on the improved MLP and CNN models,respectively.And has been effectively applied in the failure analysis of automobile engines.