Wavelet feature extraction method and system for voice signals and storage medium

Abstract

The invention discloses a wavelet feature extraction method and system for a voice signal, and a storage medium. The wavelet feature extraction method comprises the following steps: carrying out the wavelet processing of an obtained original voice signal, and obtaining a wavelet coefficient of the original voice signal; arranging the wavelet coefficients according to a first preset rule; calculating the variance of the arranged wavelet coefficients; normalizing the variance of the wavelet coefficient; arranging the normalized variances of the wavelet coefficients according to a second preset rule; and converting the variance of the arranged wavelet coefficients into a grayscale image to obtain wavelet features of the original voice signal. According to the wavelet feature extraction method, wavelet processing is carried out on the original voice signal and normalization processing is carried out on the variance of the wavelet coefficient, so that the influence of noise on the feature extraction process is reduced and thus the anti-noise capability during the feature extraction process of the non-stationary voice signal and the overlapped voice signal is improved. The wavelet feature extraction method can be widely applied to the technical field of speech recognition.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a wavelet feature extraction method, system and storage medium for speech signals. Background technique [0002] The speech recognition system will be affected to varying degrees in different environments. For example, the recognition accuracy rate is higher in a quieter laboratory environment, but the recognition accuracy rate is lower in a noisy road environment. In different noise environments, the speech recognition system will have different recognition effects, so improving the anti-noise ability of the speech recognition system is the key to improving the recognition rate and robustness of the speech recognition system. [0003] In the prior art book, the anti-noise ability of the speech recognition system is improved mainly through two aspects. The first aspect is to improve the feature extraction method of speech, such as PLP and MFCC feature extraction method...

AI Technical Summary

Problems solved by technology

[0004] In the existing feature extraction methods, such as PLP and MFCC, they all use the smooth signal processing method for filtering, denoising and feature extraction. However, this type of processing method has certain limitations: it cannot achieve a good suppression effect on abrupt noise, and the one-dimensional Fourier transform is used in the processing process, because the speech signal is non-stationary For this kind of non-stationary signal, the Fourier transform cannot achieve better results, and when the noise spectrum and the speech signal spectrum overlap too much, the Fourier transform cannot remove the noise. For some more complex environments, these Feature extraction methods are even more powerless

Images