Transferring Age and Gender Attributes for Dimensional Emotion Prediction from Big Speech Data Using Hierarchical Deep Learning

The continuous speech emotion prediction is a challenging task. This research has many important applications in real-life, especially on human-computer interaction and its applications are expected to become increasingly wider and deeper. Previous studies have show that three dimensional attributes of emotion such as arousal, valence and dominance are related each other. Moreover, age and gender of speaker affect the recognition of speech emotion. Based on these observation, we propose a new dimensional attributes prediction model using multi-task learning, aiming to improve the emotion recognition performance and generalization capabilities. The present methods utilize the big speech data to train the age and gender submodel, which will be transferred to the main modelłła hierarchical deep learning model, using age and gender as the high level attributes of the emotion. The publicly available databases IEMOCAP and aGender have been conducted to evaluate the performance and the accuracy of the proposed work. Experiment results of within-corpus evaluation show that the proposed approach has superior performance compared to state of the art result.