Unsupervised Stress Information Labeling Using Gaussian Process Latent Variable Model for Statistical Speech Synthesis

In Thai language, stress is an important prosodic feature that not only affects naturalness but also has a crucial role in meaning of phrase-level utterance. It is seen that a speech synthesis model that is trained with lack of stress and phrase-level information causes incorrect tones and ambiguity in meaning of synthetic speech. Our previous work has shown that manually annotated stress information improves naturalness of synthetic speech. However, a high time consumption is a drawback of the manual annotation. In this paper, we utilize an unsupervised learning technique called Bayesian Gaussian process latent variable model (Bayesian GP-LVM) to automatically put stress annotation on the given training data. Stress related features are projected onto a latent space in which syllables are easier classified into stressed/unstressed classes. We use the stressed/unstressed information as an additional context in GPR-based speech synthesis. Experimental results show that the proposed technique improves naturalness of synthetic speech as well as accuracy of stressed/unstressed classification. Moreover, the proposed technique enables us to avoid ambiguity in meaning of synthetic speech by providing intended stress position into context label sequence to be synthesized.