Feature Extraction And Recognition Of Sound Events In Noise Background

In recent years,the sound recognition technology develops rapidly.The sound events in the living environment are closely related to people's lives.The sound events recognition technology has certain value in the field of sound tracking,positioning and security.The recognition of sound events can reveal the events that occur in the scene and can be effectively monitored for specific locations.Due to the complex background noise in the real environment,the poor recognition performance of the recognition system in the noise background has become a major problem.The sound events recognition system is composed of sound signal preprocessing,feature extraction and classification.Feature extraction is an important link,and feature extraction directly affects the recognition effect.there are more information of extracting the spectrogram features than the traditional sound features in the case of low SNR,and the characteristics of the spectrogram are studied in depth and an improved algorithm is proposed,combined with the improved sound enhancement algorithm in the front-end processing.The main contents of the paper are as follows:(1)The sound enhancement.Comparing the traditional sound enhancement algorithms,and researching the shortcomings of Wiener filtering,multi-band spectral subtraction and other algorithms.The Improved Minimum Mean Squared Error algorithm is proposed and the Improved Minimum Constrained Recursive Averaging method is introduced.the algorithm can improve the signal-to-noise ratio of the noisy tone signal to achieve the sound signal enhancement effect,thereby improving the sound event recognition performance.(2)Based on the feature extraction of spectrogram.The basic idea of the algorithm is to convert the sound signal into a sound spectrum by using a Gammatone filter,and extract the sub-band power distribution of the sound spectrum,and the sound spectrum can include time domain and frequency.The information of the domain,the Sub-band Power Distribution contains the power distribution of different frequency bands,which can distinguish the noise and sound events better.The first layer features the global features of the spectrogram,and the second layer features the fan-shaped projection method.The detailed features of the spectrogram are combined with the Spectrogram Fan projection algorithm of the spectrum.The combination of the two features a certain anti-noise performance,which can effectively improve theperformance of sound event recognition.(3)Recognition.According to the characteristics of different classifiers,the classification and recognition experiments are carried out respectively,and a better identification framework is obtained.Finally,the support vector machine recognition framework is used to classify and identify the extracted feature matrix,and find a hyper-plane.To segment the data to achieve the classification effect,use the multi-classification method to establish the classification model.For the case of less data samples,the support vector machine can achieve better recognition results.For the recognition of sound events at low SNR,the paper classifies and recognizes 16 different types of sounds in daily environment under the conditions of speech noise,car noise,factory noise,pink noise,and different signal-to-noise ratios.Experiments show that Combined with the Improved Minimum Mean Squared Error sound enhancement algorithm and Spectrogram Fan Projection features,the real-world sound events can be identified,and a good recognition rate can be achieved at low SNR.