基于BERT-CRF模型的缅甸语韵律单元边界预测

Prosodic Unit Boundary Prediction of Myanmar Based on BERT-CRF Model

摘要: 近年来，缅甸语语音合成引起了众多学者的关注，然而该技术的性能离推广应用还有一段距离。本文以提升缅甸语语音合成自然度作为目标，研究缅甸语韵律特征，探索通过缅甸语文本自动预测韵律单元边界的方法。本文提出并实现了一种基于BERT预训练模型和条件随机场(CRF)模型相结合的缅甸语韵律词和韵律短语边界预测方法。实验结果表明，采用BERT-CRF模型，韵律词和韵律短语的预测效果均优于CRF、BiLSTM、BiLSTM-CRF以及BERT模型。为了验证该方法的可用性，本文还将本文所提出的方法应用于语音合成前端文本分析与处理中。语音合成实验结果表明，本文所提方法能有效提高缅甸语语音合成的自然度。

Abstract: In recent years, Myanmar speech synthesis has attracted the attention of many scholars, but the performance of this technology is still a long way from popularization and application. In order to improve the naturalness of Myanmar speech synthesis, this paper studies the prosodic features of Myanmar language and explores the method of automatically predicting the boundary of prosodic units through texts. In this paper, a boundary prediction method of prosodic words and phrases in Myanmar language based on BERT pretraining model and Conditional Random Field (CRF) model is proposed and implemented. The experimental results show that the prediction effect of prosodic words and phrases using BERT-CRF model is better than that of CRF, BiLSTM, BiLSTM-CRF and BERT models. In order to verify the availability of this method, the method proposed in this paper is also applied to the front-end text analysis and processing of speech synthesis. The experimental results of speech synthesis show that the proposed method can effectively improve the naturalness of Myanmar speech synthesis.