ENF-GRAD: diffusion neural vocoder based on envelope features with data adaptive prior
Ziqi Zhang, Bowen Zhang, Jian Ming Zhao, Zhonglin Jiang, Yong Chen, Chunhui Wang, Siyu Zhang · 2025
In this paper, we propose a neural diffusion-based neural vocoder model with the goal of generating high-quality realistic speech signals, named EnfGrad. The speech quality of synthesized speech is improved by introducing envelope features, which is extracted from the fundamental frequency of speech. And we reduce the number of iterations in the sampling process by using noise with a prior information to generate signals faster. By introducing dilated convolutions, temporal receptive field of the model is increased without losing information. Compared to other models, EnfGrad has a large improvement in objective metrics while maintaining approximately the same speed of inference.