Joseph Turian

We want to sample SoftModeSelector params in a way that maximize KL divergence from uniform. https://stats.stackexchange.com/questions/514415/scaling-0-1-values-so-that-their-l1-norm-is-not-uniform

Can we create a sine lookup table? or https://forums.developer.nvidia.com/t/a-faster-and-more-accurate-implementation-of-sincosf/44620 etc. Profiling

Make sure batch size is fine for consumer GPU memory with fp16 (or if it doesn't work)

3

Need to pick the batch size so people can do something useful with it downstream I guess

Generate 10 sounds. Compute the STFT between them and validate they are as expected?

Could make things faster, but a faff to set up

Just for the tutorial and futzing around

Release models or provide small librispeech recipe?

4

Since the VCC data is not commonly available could you either: * Release your pretrained models or * Add a VCTK / LibriSpeech recipe? (Since that data is available freely)

In the documentation, do you mind providing some tips on what advanced features to use to achieve the highest-quality resampling, regardless of speed?

`ddsp.spectral_ops.reset_crepe()` should be documented

1

In calling the model multiple times to create audio longer than 4 seconds, I ran into a GPU OOM issue caused by crepe. `ddsp.spectral_ops.reset_crepe()` is used in the notebooks, but...