Expressive TTS Training With Frame and Style Reconstruction Loss

We propose a novel training strategy for Tacotron-based text-to-speech (TTS) system that improves the speech styling at utterance level. One of key challenges in prosody modeling is lack reference makes explicit difficult. The proposed technique doesn’t require annotations from data. It attempt to model explicitly either, but rather encodes association between input text and its styles using TTS framework. This study marks departure style token paradigm where modeled by bank embeddings. adopts combination two objective functions: 1) frame level reconstruction loss, calculated synthesized target spectral features; 2) deep features speech. loss formulated as perceptual ensure taken into consideration during training. Experiments show achieves remarkable performance outperforms state-of-the-art baseline both naturalness expressiveness. To our best knowledge, this first incorporate quality function Tacotron improved