RESEARCH ON THE EFFECT OF A MICROPHONE SELF-NOISE SUPPRESSION BY USING VARIOUS METHODS FOR AN NOISED AUDIO SIGNAL FOR CLASSIFICATION

Abstract

The paper presents a comparative analysis of noise reduction methods: wavelet transformation, adaptive RLS filtration, and spectral subtraction methods (by amplitude and power). The experiment was performed by modeling an additive mixture of an ideal signal, which is taken from the ESC-50 dataset and belongs to the “engine” category, and white noise with a fixed signal-to-noise ratio (SNR) at the level of 16.51 dB. The performance of a signal denoising was measured by using two criteria: the increase in the SNR metric and the relative reduction in the mean square error (MSE), calculated relative to the noisy signal. The modeling results showed that the best denoising performance is provided by the RLS filter, which increased the SNR by 11.51 dB. The wavelet transformation method showed an improvement of 5.09 dB, while the spectral subtraction methods in terms of power and amplitude provided an increase of 5.2 dB and 3.75 dB, respectively. The experimental modeling showed that the white noise reduces the recognition accuracy from the idel 90.61% to 62.57% for the classifier (trained on the ESC-50 dataset). The usage of the RLS filter allowed to improve the accuracy to 90.16%, which is the best result among the considered methods. The wavelet transform (with aggressive noise suppression settings) provided an accuracy of 80.61%. The spectral subtraction methods turned out to be less effective for the classification task, showing a result of 65.02% (in power) and 64.7% (in amplitude).