yangdongchao / AcademiCodec

AcademiCodec: An Open Source Audio Codec Model for Academic Research
574 stars 80 forks source link

fix #33 #38

Closed yt605155624 closed 1 year ago

yt605155624 commented 1 year ago

using more agressive discriminator (6 difference complex domain STFT)

fix #33

在 Encodec_24k_240d 上用 CSMSC (数据集提前下采样到 24k) 可以训练起来

image image

discriminator_iter_start 修改为 1 后也可以正常打印 d_loss

image

遗留问题参考