Speech-Backbones

Why we included this project

Huawei Noah's Ark Lab's speech work is spread across three papers, and this repo gathers the official code for all of them in one place. Grad-TTS is a diffusion-based text-to-speech model that builds mel-spectrograms by denoising step by step, so it is a handy reference if you are exploring probabilistic TTS. SPIRAL learns speech representations from unlabeled audio without needing labels, which makes it a reasonable pre-training option when you want to build recognition or understanding systems from raw recordings. DiffVC applies the same diffusion idea to voice conversion, with a fast maximum-likelihood sampling scheme, and is a good place to start for identity-transfer experiments. Each sub-project links to its paper and includes the code, so the repo works as both a study resource and a base to adapt for your own synthesis, pre-training, or conversion work.