unconditional synthesis

Open Rongjiehuang opened this issue 3 years ago • 1 comments

Hi, thanks for sharing this code with the community. Before this, I tried several repos but failed to get reasonable sound in the SC09 dataset.

I am excited to generate waveforms with your provided checkpoint (model=wavenet model.res_channels=128 model.num_res_layers=30 model.dilation_cycle=10 train.batch_size_per_gpu=4) and get samples. However, when I conduct testing following this, I fail to reproduce the objective evaluation results in the original diffwave paper (mIS: 5 vs. 59), and some items are noisy.

Have you ever met with this issue? and Could you recommend a code base for unconditional waveform synthesis? Thanks!

Jul 23 '22 08:07 Rongjiehuang

That command corresponds to the small DiffWave(+WaveNet) model, which is not supposed to sound great. Instructions for the base model are in the README. Our paper reports mIS ~9 for the small model and ~51 for the base model (Table 7).

I'm not aware of another codebase that has good unconditional SC09 synthesis.

Jul 23 '22 15:07 albertfgu