Research On Emotion Recognition Of Speech Signal Based On HMM

With the popularity of computers and the rapid development of computer science and technology, the dependence of human on computer is getting heavy. As a result, more and more people began to focus on the new interactive technology. Language, as a human-specific feature, plays an important role in exchange of people, such as exchange of ideas, opinions and emotional communication, etc. Hence, more attention has to be paid to the research on emotional information affiliated to Speech Signal. As an important branch of Emotional Speech Signal Processing, emotional speech recognition plays a key role in human-computer interaction.This paper firstly introduces the research background, significance, current research situation, and its application prospect. The following part presents emotion classification and common methods for emotion recognition at home and abroad. On the basis of domestic and abroad emotional speech database, an emotional speech bank including "happiness", "peace", "anger" is recorded in terms of emotional recognition, test environment, equipment requirements and etc. By subjective listening recognition tests, speeches with high-recognition are selected for the emotional speech recognition test. After an analysis on emotional characteristics of the speech signal, methods for various parameters extraction used in the test of this paper are described in detail. Parameters extracted are as follows:Pitch frequency, the first and second derivatives of Pitch frequency, the first, second and third derivatives of short-term energy, the first resonant with its traditional characteristic parameters and anti-noise characteristic parameters, etc. The modeling thoughts and methods for emotional speech recognition model are also especially introduced in this paper. As a research method, the continuous Hidden Markov Model (CHMM) with a jump from left to right is used in emotional speech recognition in this paper. The whole process of this research is composed of model initialization, parameters training and emotion recognition. In the emotion recognition training, single MFCC characteristic parameters, single ZCPA characteristic parameters and nine dimension of emotion characteristic parameters vector are respectively tested in calm and noise situations. Experimental results show that the results of three kinds of emotion recognition are ideal. In particular, the "angry" has a better recognition rate compared with the other two emotions.under the same experimental conditions. Recognition result of single ZCPA characteristic parameters and vectors for comprehensive emotional characteristics parameters are better than the result of the traditional MFCC feature parameters. Finally, some existing problems in this field, the insufficiency of this research and the further research direction are discussed in this paper.