Network Compression Using Specialists

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

Ground-truth speech source

Recovered speech from
the proposed specialist (512X2)

12.86 dB SiSDR

Noise source

Recovered speech from
the small generalist (512X2)

11.31 dB SiSDR

Mixture (-5.33 dB SiSDR)

Recovered speech from
the large generalist (1024X3)

12.99 SiSDR

Example #2

Ground-truth speech source

Recovered speech from
the proposed specialist (512X2)

11.37 dB SiSDR

Noise source

Recovered speech from
the small generalist (512X2)

10.21 dB SiSDR

Mixture (-4.36 dB SiSDR)

Recovered speech from
the large generalist (1024X3)

11.13 SiSDR

Source code

https://github.com/IU-SAIGE/sparse_mle