VALL-E 1.0: Reproduce the results of the paper, Train larger models, and greatly improve the speed of token extraction.

VALL-E 1.0 Release notes

• Reproducted Demo https://lifeiteng.github.io/valle/index.html
  • Traning commands egs/libritts#demo
  • Inference valle/run.sh#L68
• Support Two stages training #59
• Support Mandarin #85 Thanks @zhaomingwork
• Speedup Encodec AudioTokenizer #81 #88 Thanks @Guokr233
• Thanks reviewer @zjwang21

Greatly improved accuracy and Fix GPU memory increase during training

• improved AR Top10 accuracy use default optim scheduler and base_lr
• Fix GPU memory increase during training

Train VALL-E on a single GPU!

v0.1.0

v0.1.0