Research On High-performance Audio Scene Recognition Technology

From a single natural event to the overall sound field,the sounds contain a variety of information that humans use to understand the surroundings.In recent years,several novel methods for automatically analyzing such information have been proposed,and some new applications have emerged.By analyzing audio signal,audio scene recognition can extract the environmental information,and obtain the essence of audio content,which improve human life and provide better serve for human society.In addition,some related key technologies of audio scene recognition,such as signal processing,data mining and data decision,play an important role in analyzing nonlinear non-stationary signals,intelligent multimedia information processing and retrieval,promoting the integration of information science and mathematics,and promoting the development of new technologies such as artificial intelligence.Audio scene recognition has been widely applied to the description and retrieval of audio and video,machine navigation,speech noise reduction,scene modeling,the intelligence acquisition of military criminal investigation,and monitoring,et al.This article aims at the high performance of audio scene recognition,and mainly completed the following work:(1)Analysis the principle of audio scene recognition.Based on the theory of pattern recognition,the audio scene recognition is analyzed in detail,which provides a theoretical basis for audio scene recognition.Besides,based on the general model of audio scene recognition,an improved audio scene recognition model with feature selection algorithm is proposed to efficiently build the model of the audio scene.(2)Study the audio scene feature acquisition method.A series of features of the audio scene signal are extracted.For the problem of computational complexity and low performance in the existing feature selection algorithm,an improved optimization framework for feature selection is proposed.Based on this framework,the feature selection algorithm based on an improved feature evaluation criteria and shuffled leap algorithm is proposed to obtain an optimal feature subset of the audio signal feature and improve the effectiveness of the audio scene feature.(3)Study the accurate evaluation criteria of classifier performance.For the multi-classification task,a comprehensive evaluation criteria of the classifier based on the confusion matrix are proposed,which applied to evaluate the performance of the complex audio scene recognition system to verify the validity and advantages of the proposed method in this thesis.(4)A classification method for model fusion based on GMM and CNN is proposed to improve the classification performance of the system.(5)Design and implement system.Based on GMM and CNN,the audio scene recognition system is built.The experimental results of the audio scene recognition system implemented in this paper can quickly and accurately classify these audio scenes,which has the advantages of high accuracy,low cost and simple operation,and research reference value.