Research On Application Of Matching Pursuit In Environmental Acoustic Event Recognition

Environmental sounds contain a wealth of information, and can provide efficacious data support for nature conservation, apply to robot navigation, home automation as well as mobile terminal equipment. Research on the recognition method of environmental acoustic events in real-world has a more direct theoretical significance and practical value. However, performance of the recognition system would have a sharp decline in the noisy environment, because a large amount of background noise exists. The training environment, therefor, cannot match that. In consider of this, this paper discussed from the two aspects:improve quality of the input samples and extract robust features, and moreover, studied the application of matching pursuit algorithm in environmental sound events recognition in actual environments.The main work includes the followings:1) Establish the environmental sound events database. The database has 5 categories, which add up to 45 kinds of signals, including sounds of animals, birds, insects, the nature and sounds of everyday object. All 45 kinds of signals were analyzed with their fundamental frequency and thus laid the foundation for further study.2) MP with the Gabor wavelet extracts 25 Gabor atoms from the entire audio to achieve the purpose of eliminating the influence of background noise. The signal is reconstructed from the extracted atoms. Moreover, using the genetic algorithm to optimize the rate of decomposition of the MP is studied.3) Extract frame features of MP/Gabor in order to reduce the recognition parameters as well as enhance the extracted features’ robustness. Considering that the traditional Mel frequency cepstral coefficients (MFCCs) are noise sensitivity, the frequency domain and time-frequency domain frame features, fundamental frequency, MFCCs-12, MFCCs-26, MFCCs-39 and MP-4, are extracted. When calculating the MP feature, decomposing speed of MP has to be taken into account. Finally, we obtain the best atomic parameters reconstructed by 7 atoms.4) Classification algorithm. According to the above step, we construct the Support Vector Machine (SVM) to accomplish the recognition task. By the way of comparative experiment, we present the classification results before and after MP and different features as well. Considering that in database the audio signals’ signal-to-noise ratios (SNR) vary, Gaussian white noise with various SNR is used as background noise to test the validity of MP for noise reduction.In this paper, environmental sound events of 5 categories are studied. Then the comparison experiments are designed and conducted. The experimental results show that MP sparse representation of the entire signal can eliminate the influence of background noise effectively, but it cannot meet the real-time requirements; the feature set (fundamental frequency, MFCCs-12 and MP-4) is able to characterize the environmental acoustic events in real-life situation, has a better average recognition rate and thus is noise robustness and can provide a reference for further study of environmental acoustic sounds in real life scenes.