Voice recognition method based on glottis wave information

Abstract

The invention discloses a voice recognition method based on the glottis wave information. The method comprises three steps of source signal preprocessing, feature extraction, and classification and recognition, wherein the source signal preprocessing extracts a glottis wave signal of an original voice signal as a source signal for feature extraction; the feature extraction adopts a dynamic image expert group standard MPEG-7 to extract an audio high-order statistic feature, and combines the audio high-order statistic feature with an openSMILE feature and a classic glottis feature to serve as a feature set for glottis wave signal identification; classification and recognition are based on a random forest classifier, and prediction and classification of speech recognition are carried out by adopting a ten-time cross validation method. According to the method, glottis waves are used as source signals, the effects of glottis excitation and vocal cord vibration mechanisms in speech recognition are fully represented, and a dynamic image expert group standard MPEG-7 is proposed to extract audio high-order statistical characteristics and combine the audio high-order statistical characteristics with openSMILE characteristics and classic glottis characteristics to serve as a recognition characteristic set; a problem of inter-frame repetition and over-fitting is solved, and meanwhile, the method does not depend on a fundamental tone frequency estimation result.

Description

technical field [0001] The invention relates to the field of voice recognition, in particular to a voice recognition method based on glottal wave information. Background technique [0002] Speech recognition technology can generally be decomposed into stages such as signal preprocessing, feature extraction, and classification recognition. The feature categories extracted by speech recognition technology can be divided according to their principles: measurement disturbance based on pitch frequency and signal amplitude, cepstrum obtained by spectral cepstrum calculation, nonlinear dynamic analysis and calculation, and Glottal class estimated by inverse filtering algorithm, etc. Among them, the perturbation features and nonlinear features are subject to the estimation accuracy of the pitch frequency, and their performance is not as good as that of the cepstrum features. The glottal features are mostly used as supplementary features, and their value in speech recognition has no...

AI Technical Summary

Problems solved by technology

The consideration of speech recognition performance improvement in most technical solutions is limited to which feature performs better and which classification algorithm performs better. These solutions use the original speech signal as its source signal for analysis and feature extraction, and cannot eliminate Due to the influence of lip radiation and vocal tract resonance, it is difficult to highlight the important characteristics of glottal excitation generated by vocal cord vibration in the process of vocalization

Some technical solutions use glottal features in speech recognition as supplementary features to improve the feature set to describe the integrity of speech signal characteristics. However, the main body is still mostly cepstrum, nonlinear or measurement disturbance features. The effect of speech recognition performance improvement has not yet been reflected. In summary, the speech recognition method in the prior art has the following defects:

[0004](1) The features extracted by existing speech recognition methods are limited by the accurate estimation of pitch frequency, and the value of glottal wave information in speech recognition cannot be fully utilized , has limited contribution to improving the accuracy of speech recognition

[0005](2) The existing speech recognition method uses the original speech signal after vocal tract resonance and lip radiation as the source signal processing, which fails to eliminate vocal tract resonance and lip radiation The influence of the influence, can not fully describe the mechanism and characteristics of vocal fold vibration

[0006](3) Part of the speech recognition method adopts the feature extraction method with the frame as the analysis unit, which inevitably has the repetition and overfitting of the feature extraction of the frame shift part, which reduces the Robustness and reliability of recognition results

Images