High-Quality Text-to-Speech Implementation via Active Shallow Diffusion Mechanism.

Area of Science:

• Speech Synthesis
• Deep Learning
• Probabilistic Modeling

Background:

• Denoising diffusion probabilistic models (DDPMs) show promise for text-to-speech (TTS).
• Traditional diffusion models face challenges in real-time applications due to extensive sampling requirements.

Purpose of the Study:

• To propose a novel, efficient, and fast-inference diffusion-based TTS model.
• To enable real-time TTS synthesis using diffusion models.

Main Methods:

• Introduced Cascaded MixGAN-TTS (CMG-TTS), a two-stage diffusion model.
• Employed an active shallow diffusion mechanism for staged training.
• Utilized a mixture combination mechanism linguistic encoder with pitch and energy predictors.
• Incorporated a post-net for mel-spectrogram reconstruction optimization.

Main Results:

• CMG-TTS achieved satisfactory subjective and objective evaluation metrics with only one denoising step.
• Demonstrated leading performance in the real-time factor (RTF) compared to other diffusion-based TTS models.
• Ablation studies confirmed the effectiveness of both stages within the CMG-TTS architecture.

Conclusions:

• CMG-TTS offers an efficient and fast solution for diffusion-based TTS.
• The proposed two-stage approach significantly enhances inference speed for diffusion models in TTS.
• CMG-TTS represents a significant advancement in achieving real-time, high-quality speech synthesis.