Based On The Active Learning Of Chinese Prosodic Phrases Prediction Research

With the rapid development of computer intelligence, speech synthesis technology becomes more widely applied to people’s lives. Speech synthesis naturalness is an important direction of development of speech synthesis technology, and the accuracy of the prosody structure prediction is not high, which is a major obstacle to improve the naturalness of synthetic speech.Based on the history and current research situation of the Chinese prosodic structure prediction, this paper summarizes and analyzes the current technology used in Chinese prosodic structure prediction. Similar to other machine learning methods, the prediction of Chinese prosodic phrase requires a certain amount of labeled data as training data, so that you need to spend a lot of time and manpower on the labeling of prosodic phrase, and this is the common problems to traditional classification algorithms. In recent years, there is a new method which integrates the labeled data and the unlabeled data when training data, to overcome the lack of labeled data and to greatly reduce the workload of the manual labeling. Semi-supervised learning method and active learning method were both generated based on this idea.In this paper, the author uses an active learning method based on CRFs and realizes a prototype system for the prediction and annotation of Chinese prosodic phrase boundaries.The prototype system in this paper uses human-computer interaction, beginning with partial labeled training data, making the computer select the best sample to ask the annotator to label and update the initial training set. Experiments on the active learning method based on CRFs prove that the method can effectively solve the active learning problem in Chinese prosodic phrase predicting and labeling without so much labeled data, and the workload of manual labeling will be greatly reduced.