HierTTS: Expressive End-to-End Text-to-Waveform Using a Multi-Scale Hierarchical Variational Auto-Encoder

End-to-end text-to-speech (TTS) models that directly generate waveforms from text are gaining popularity. However, existing end-to-end still not natural enough in their prosodic expressiveness. Additionally, previous studies on improving the expressiveness of TTS have mainly focused acoustic models. There is a lack research enhancing an framework. Therefore, we propose HierTTS, highly expressive text-to-waveform generation model. It deeply couples hierarchical properties speech with variational auto-encoders and multi-scale latent variables, at frame, phone, subword, word, sentence levels. The encoder encodes signal fine-grained features into coarse-grained variables. In contrast, decoder generates conditioned We staged KL-weighted annealing strategy to prevent posterior collapse. Furthermore, employ extract linguistic information different levels act both decoder. Experiments show our model performs closer prosody has better generative diversity.