Speech Emotion Recognition Research Based On Manifold Learning And D-S Evidence Theory

Speech emotion recognition is to analyze and detect the special emotion state from given emotion speech emotion utterance and then to ascertain the subject's specific inborn emotion, which can achieve smarter and more natural interaction between human beings and computers. In this thesis, we firstly discuss the background and then analyze the main exiting speech signal pretreatment methods, feature extraction algorithms, feature dimensionality reduction algorithms and emotion recognition algorithms. After analyzing the methods currently used by others, several improved algorithms and methods for these tasks are developed. The performances of our methods are illustrated by experiment results. The main work is described as bellows:(1) Feature dimensionality reduction method based on incremental manifold learning is presented. There are many emotion features, including redundant features and irrespective features, so we present feature dimensionality reduction method based on incremental manifold learning in order to reduce the effect of redundant information. In this method, 101 features extracted from the voice of quality, energy, Fundamental Frequency, Formant Frequency, MFCCs and Mel Frequency Energy Dynamic Coefficient are used as initial features, and then Isomap are used to reduce the initial features dimensionality of training set. The low-dimensional features of test set are calculated by incremental learning. So the influence of redundant features is reduced, and the recognition rate is increased.(2) Multi-granularity segment method is proposed. Usually, long sentence is segment by fixed frame numbers or length ratio, and then emotion features are extracted and analysed. This segment method is simple to implement the fusion of result of speech emotion analysis. But it isn't fully take into account the integrity of the speech emotion information. In order to obtain the more complete and richer feature information, we present multi-granularity segment method. Sentence is segment by fixed length ratio and rhythm, and then speech emotion analysis and results fusion would be done.(3) Speech emotion recognition method of multi-granularity segment fusion based on D-S evidence theory is proposed. After segment finished, each part will be regarded as a separate sample to identify. An original sample will obtain several results. Information fusion technology is needed to fuse these results. So we present speech emotion recognition method of multi-granularity segment fusion based on D-S evidence theory. Several segment results which belong to the same sample are fused by D-S evidence theory to obtain this sample's result. Experiments show that, the recognition performance of this method is better, and the speech emotion recognition accuracy rate is improved effectively.(4) By combining Matlab with VC++, a prototype system of speech emotion recognition based on features dimensionality reduction by incremental manifold learning and multi-granularity segment fusion is designed, which demonstrates the effectiveness of the algorithms mentioned above.