A Robust Feature Extraction Method for Sound Signals Based on Gabor and MFCC

Abstract

Along with the development of data scale and the high complexity of sound signals, feature extraction and classification methods of sound signals have become a major research hotspot. However, the current sound signal feature extraction methods are difficult to accurately and stably provide a high-precision classification effect for the sound signal due to the complex frequency distribution and the influence of noise. Therefore, a robust feature extraction method for sound signals based on multi-scale and multi-directional Gabor filters and Mel frequency cepstral coefficient (MtWGM) was proposed. This method performs preprocessing on the signal by mixing hard threshold and soft threshold wavelet denoising. The Gabor filter is used to the combined energy spectrum of the framed signal, to achieve the effect of relatively more balanced intra-class features and more prominent inter-class features, and finally improve the noise reduction performance and classification accuracy of sound signals. The experiments are conducted out on three different sound signal datasets. Three classifiers are trained to test the effectiveness of extracted features. The experimental results show that the proposed multiple wavelet Gabor_MFCC (MtWGM) method has obtained better classification accuracy and robustness than that of MFCC.