Research On Emotion Recognition From Speech-Features And Models

Emotional recognition from speech becomes a hot topic currently, but because of different emotion features and recognition modals, and the fact that experiments are done on different emotional speech databases, which causes the results not comparable, it is difficult to discriminate the merits of the features and modals, especially the modal with global features and the dynamic modal with short-time features. Here we first analyze and select the emotional speech features which reflect the variation trend of the four emotions (happy, anger, sad, neutral), and compare results on the global modal and dynamic modal based on the same emotional speech database.1. An emotional speech database has been record. Scripts from standard TIMTT Englishspeech database are read by 46 individuals with four emotions (happy, angry, sad and neutral), each person repeats 25 sentences with the four emotions. Through perception subjective perception and evaluation experiment, 8 persons' 800 sentences are selected for our experiments.2. Through observing and analyzing, the variation trends on each emotion of the followingfeature curves: pitch, spectral information and speed, we elect and define a 23-dimentional global emotion features (pitch, resonance, speed, average energy, etc.) which are discriminative on the four emotions.3. The training and recognition algorithms of GMM is studied, the GMMs with globalemotion features are built for four emotions. Emotion recognition experiments show that, if only the 12-dimentional pitch related features are adopted, sad and can be correctly recognized than the other two emotions. After the resonance, speed, average energy are considered, the correct recognition rates are improved for the four emotions. Results also show that speed and average energy are discriminant for the four emotions, while resonance is useful for the distinguishing happy and angry.4. The training and (?)ecognition algorithm of HMM is studied, emotion HMMs are builtrespectively with MFCC features (feature 1), and with dynamic features including short-time energy, resonance, sub-band energy (feature 2). Emotion recognition experiments results show that, feature 2 gets the improvement of 29.55% on the...