Low-Resource Mandarin Prosodic Structure Prediction Using Self-Training

For Chinese text-to-speech (TTS) system, prosodic structure prediction plays an important role in it. The prosodic information can improve the synthesize result tremendously. Re-current neural network (RNN), conditional random fields (CRF), and self-attention have been used in this task. These approaches use large amount of data. However, the data marking process can be really time-consuming, and needs people with relevant knowledge to participate, which makes the prosodic information dataset expensive. On the other hand, the unlabeled data, are easy to get. In this paper, we use self-training to make use of unlabeled data and solve the labeled data insufficient problem. With the help of pre-trained language model, we start training with little labeled data and use active learning to generate more data. Experiments show that only starting training with 4k of labeled data, the proposed method can achieve equivalent or better result than baseline, while the baseline used more than 40k labeled data.