Apparently the AudioX and AudioLDM(2) models do this but I think you've found a genuine gap.