Improving generation performance of speech emotion recognition by denoising autoencoders

Abstract

A speech emotion recognition algorithm should generalize well when the target person's speech samples and prior knowledge about their emotional content are not included in the training data. In order to achieve this objective, we utilize denoising autoencoders based approach to solve this task. In this study, a relatively small dataset, which contains close to 1500 persons' emotion sentences, is introduced. By unsupervised pre-training with this dataset, denoising autoencoders learn features which contain more emotion-specific information than speaker-specific information in data successfully. Experiment results in CASIA dataset show that this denoising autoencoders based approach can improve the generation performance of speech emotion recognition significantly.