huawei-noah/Speech-Backbones

Implements multiple diffusion-based and self-supervised speech models: Grad-TTS for text-to-speech synthesis using diffusion probabilistic modeling, SPIRAL for speech pre-training via perturbation-invariant representation learning, and DiffVC for voice conversion with fast maximum likelihood sampling. The collection covers the full speech synthesis and processing pipeline with published research-backed implementations accepted at top-tier venues (ICML, ICLR).

602 stars. No commits in the last 6 months.