| Speech is the most natural way of communication. In order to achieve human-computer interaction and how to make the computer to understand human language has become the focus of current research direction. Speech Emotion Recognition refers to the computer through the analysis of the speaker’s emotional state and its changes to determine the speaker’s emotional feelings, making human-computer interaction more comfortable and harmonious. Speech emotion recognition technology can enhance the level of intelligence of the computer, and has a very important social value and practical significance.Many machine learning algorithms such as maximum entropy, hidden Markov, support vector machines and other models as emotion recognition classifiers have applied to speech emotion recognition, and has reached a certain level of recognition accuracy. However, most of these methods is based on a large scale corpus, artificial label of the training samples is a very time-consuming, especially for speech emotion recognition database, so far available resources is still very limited. Therefore, how to use the information of the unlabeled samples to improve the generalization ability of the classifier has become a very popular field of study.Active learning has the advantages of high learning efficiency and classification performance can be used to solve the small sample learning problems in supervised learning. Based on this, active learning is applied to speech emotion recognition, and use the information density of query strategy to select the sample.First, a Chinese emotional speech database has been constructed for experiments. Based on the selected features and analysis of algorithms used in speech emotion recognition, this paper proposes a new active learning strategy based on Information Density (ID) integrated with CRFs for Speech Emotion Recognition. And development a prototype system for speech emotion recognition based on active learning with CRFs. The experiments show that without large-scale labeled data, the proposed method greatly reduces the training time and gets better results as compared to the conventional supervised learning mode. |
