US 12,136,413 B1
Domain-specific parameter pre-fixes for tuning automatic speech recognition
Saket Dingliwal, Bellevue, WA (US); Sravan Babu Bodapati, Redmond, WA (US); Katrin Kirchhoff, Seattle, WA (US); Ankur Gandhe, Bothell, WA (US); Anubhav Mishra, Seattle, WA (US); John Baker, Bellevue, WA (US); Ashish Vishwanath Shenoy, Seattle, WA (US); and Ravi Teja Gadde, Los Angeles, CA (US)
Assigned to Amazon Technologies, Inc., Seattle, WA (US)
Filed by Amazon Technologies, Inc., Seattle, WA (US)
Filed on Mar. 31, 2022, as Appl. No. 17/710,762.
Int. Cl. G06F 40/40 (2020.01); G10L 15/06 (2013.01); G10L 15/183 (2013.01)
CPC G10L 15/063 (2013.01) [G10L 15/183 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A system, comprising:
at least one processor; and
a memory, storing program instructions that when executed by the at least one processor, cause the at least one processor to:
receive speech data via an interface for a speech processing application;
apply a speech model to the speech data to select a plurality of candidate texts predicted for the speech data by the speech model according to respective prediction scores of the plurality of candidate texts;
append domain-specific parameters as respective pre-fixes to individual ones of the plurality of candidate texts, wherein the domain-specific parameters are trained using domain-specific unlabeled text data input to a pre-trained transformer-based language model;
input the plurality of candidate texts with respectively appended domain-specific parameters to the pre-trained transformer-based language model to generate respective perplexity scores for the plurality of candidate texts; and
select one of the plurality of candidate texts for further speech processing by the speech processing application according to the respective perplexity scores of the plurality of candidate texts.