A Manifold-Based Speech Emotion Recognition Method

Abstract

The invention provides a speech emotion recognition method based on manifold. The speech emotion recognition method based on the manifold comprises the following steps that the phonetic features of a test statement are extracted, wherein the phonetic features comprise MFCC, LPCC, LFPC, ZCPA, PLP and RASTA-PLP; the local average value and the variance of each extracted phonetic feature are calculated, the local average value and the variance of the first difference of each extracted phonetic feature are calculated, the local average values and the variances are connected in series, and therefore local statistical features of the test statement are formed; a universal background model (UBM) and the local statistical features of the test statement are applied so that a specific Gaussian mixture model (GMM) of the test statement can be generated, then, all average values of the GMM are connected into a vector, and the vector serves as a feature vector of the test statement; according to the features selected according to the integrated feature selection algorithm and the multi-cluster feature selection algorithm (MCFS), the feature vector of the test statement is changed; a support vector machine classification model is applied, the feature vector, existing after feature selection, of the test statement serves as input, and emotion classes of the statement are tested in a classified mode. The degree of accuracy of speech emotion recognition of the speech emotion recognition method based on the manifold is high.

Description

technical field [0001] The invention relates to the field of speech signal processing and recognition, in particular to a manifold-based speech emotion recognition method. Background technique [0002] With the continuous development of information technology, people put forward higher requirements for the intelligence of computers. In terms of human-computer interaction, a computer with emotion recognition capabilities can recognize human emotions and perform corresponding actions based on the recognition results, which can facilitate users to operate the device and bring users a good experience. For example, the use of emotion recognition technology can detect whether the driver's energy is concentrated, the level of pressure he feels, etc., and decide whether to issue an alarm based on the recognition results to improve driving safety; at the same time, emotion recognition can also be applied to robots, smart toys, games, In e-commerce and other related industries, it he...

AI Technical Summary

Problems solved by technology

[0006] The first method: generating statistical features through global statistics, its disadvantage is that it ignores the local information of the speech signal

[0007] The second type: only generate statistical features through the Gaussian Mixture Model-Universal Background Model GMM-UBM (GaussianMixtureModel-UniversalBackgroundModel). Although this method can reflect the information of each frame of speech, it ignores the relationship between adjacent frames, and The local information extracted by it is reflected on each frame, and the granularity is too fine

These feature selection methods all select or remove a feature to minimize some objective equations, without considering the internal structure of high-dimensional speech data, and it is difficult to select a suitable feature subset in high-dimensional speech data. These limitations reduce speech Accuracy of emotion recognition

Images