| Environmental sound recognition refers to the process of allowing computer equip-ment to classify and process the collected environmental sounds in a certain algorithm.It is used in intelligent robots,mobile device monitoring,autonomous driving,environ-mental safety monitoring,smart homes,smart cities and other fields.It has a wide range of application prospects and is one of the important research directions in the field of Computer Audition.Environmental sound classification and enhancement are important research contents in the field of sound recognition.With the continuous development of artificial intelli-gence technology,neural networks have gradually become an important research method in this field.Compared with traditional machine learning methods,neural network meth-ods have the advantages of simple algorithm model,good generalization effect,strong robustness,and strong transferability.However,classification algorithms still have the problems of low recognition accuracy and high requirements for equipment;enhance-ment algorithms are also mainly oriented to scenarios such as speech enhancement and musical instrument sound separation,and environmental sound enhancement algorithms are still immature.Focusing on the above problems,this dissert has the following main research con-tents:The bottleneck module is used to improve CNN to reduce the amount of network parameters and calculations;the SE attention mechanism model is introduced to improve network accuracy.In the dissert,BN-CNN,SE-CNN small and SE-CNN large neural network models for environmental sound classification are designed.Then,different enhancement networks are selected for different classification results to perform sound enhancement processing.Based on the method of generating a con-frontation network,this dissert conducts a sound enhancement algorithm research.The performance of the generation network is improved by introducing the encoder-decoder direct connection structure;the robustness of the network is improved by introducing the noise vector z.The network can be retrained to achieve the expansion of other types of environmental sound enhancement.Next,because the current open source sound data set is not suitable for environmental sound enhancement,the dissert created the ESCS data set.This data set contains two sub-data sets,which are respectively used for the algorithm research of sound classification(12 sub-categories,5 major categories)and sound enhancement(10 sub-categories,including noisy and non-noise versions),a total of 14.6 Thousands of environmental sound samples with 44.1KHz sampling,8bit quantization,and 8s duration.At the same time,the original data provided by the data set can be processed as needed.In this data set,the accuracy of fine classification F1 of sound classification is increased by up to 4.90%,the accuracy of rough classification F1 is increased by up to 3.46%,the amount of calculation is reduced by up to 26.08%,and the amount of parameters is reduced by up to 79.05%;the Generator and the Discriminator are used to compete against each other.According to the method,the FID index is 0.218 for ten different environmental sounds;the highest average value of the AMT index is 8.12,all of which have obtained good results of environmental sound enhancement.Finally,a visualized environmental sound recognition software is designed and im-plemented.The software uses python libraries such as tkinter,librosa,and pyaudio,and is composed of an input module,a signal preprocessing module,a sound recognition module,a sound enhancement module,and an output display module,which realizes the functions of environmental sound classification and sound enhancement. |
