US 11,676,006 B2
Universal acoustic modeling using neural mixture models
Amit Das, Redmond, WA (US); Jinyu Li, Redmond, WA (US); Changliang Liu, Bothell, WA (US); and Yifan Gong, Sammamish, WA (US)
Assigned to Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on May 16, 2019, as Appl. No. 16/414,378.
Claims priority of provisional application 62/834,569, filed on Apr. 16, 2019.
Prior Publication US 2020/0334527 A1, Oct. 22, 2020
Int. Cl. G06N 3/08 (2023.01); G10L 15/16 (2006.01); G10L 15/22 (2006.01); G06N 3/04 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/04 (2013.01); G10L 15/16 (2013.01); G10L 15/22 (2013.01)] 21 Claims
OG exemplary drawing
 
1. A universal modeling system that performs domain expert model mixing, said universal modeling system comprising:
a plurality of domain expert models, each of which receives raw input data and generates a corresponding domain expert output based on the raw input data such that multiple domain expert outputs are generated;
a neural mixture component that receives an input of a stacked vector comprising a stacking of the multiple domain expert outputs and that generates a respective weight corresponding to each one of said domain expert models such that multiple weights are generated, said multiple weights being generated based on the multiple domain expert outputs that were received from and generated by the plurality of domain expert models; and
an output layer that provides a universal modeling system output, the universal modeling system output is generated based on each domain expert's corresponding domain expert output and that same domain expert's corresponding weight, which was generated based on the multiple domain expert outputs.