For the paper, “Sparse Mixture of Local Experts for Efficient Speech Enhancement,” submitted to Interspeech 2020.
The point here is that our proposed specialist network, although it’s less complex (512 X 2), can compete with a large generalist (1024 X 3).
Example #1


the proposed specialist (512X2)


the small generalist (512X2)


the large generalist (1024X3)
Example #2


the proposed specialist (512X2)


the small generalist (512X2)


the large generalist (1024X3)