ASLP-lab

/

DiffRhythm-base

like 136

Echo that question

We'll be releasing the training code on github soon.
We support a variety of styles, you can use any reference audio to provide style information

Reference audio does not really keeps the generated audio close, does not work that well, also the reelased weights do not understand basic structure of the music - verse, chorus etc. It generates pretty random audio without coherent structure, is the bigger 4minute weights version better in this aspect ? Audio quality is good but withouyt basic structure it does not have much use , too random.