Investigating deep neural network adaptation for generating exclamatory and interrogative speech in Mandarin

Abstract

Currently, most speech synthesis systems generate speech only in a reading style, which greatly affects the expressiveness of synthetized speech. To improve the expressiveness of synthetized speech, this paper focuses on the generation of exclamatory and interrogative speech for Mandarin spoken language. We propose a multi-style (exclamatory and interrogative) deep neural network-based acoustic model with a style-specific layer (can have multiple layers) while other layers are shared. The style-specific layer is used to model the distinct style specific patterns. The shared layers allow maximum knowledge sharing between the declarative and multi-style speech. Such method is also validated on different neural networks. Besides, models that adapted on both spectral and F0 parameter is compared with models while only F0 parameter is adapted. Both subjective and objective evaluations show that this method is superior to prior work (which is trained by the combination of constrained Maximum likelihood linear regression (CMLLR) and structural maximum a posterior (SMAP)), and the BLSTM where top 1 layer is style-specific layer that adapted on both spectral and F0 parameter can achieve the best results.