US 12,067,978 B2
Methods and systems for confusion reduction for compressed acoustic models
Fuliang Weng, Itasca, IL (US); Alexei Ivanov, Itasca, IL (US); and Stephen Cradock, Itasca, IL (US)
Assigned to Samsung Electronics Co., Ltd., Suwon-si (KR)
Filed by Samsung Electronics Co., Ltd., Suwon-si (KR)
Filed on Jun. 1, 2021, as Appl. No. 17/335,663.
Claims priority of provisional application 63/033,434, filed on Jun. 2, 2020.
Prior Publication US 2021/0375270 A1, Dec. 2, 2021
Int. Cl. G10L 15/065 (2013.01); G06F 40/237 (2020.01); G10L 15/187 (2013.01); G10L 15/22 (2006.01)
CPC G10L 15/187 (2013.01) [G06F 40/237 (2020.01); G10L 15/22 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A user device, comprising:
a processing circuit;
an acoustic engine;
a decoder; and
memory storing instructions that, when executed by the processing circuit, cause the user device to:
expand a speech lexicon of the decoder with recovered grapheme or phoneme sequences corresponding to a first command and a second command represented in the speech lexicon;
renormalize weights of the speech lexicon corresponding to the first command and the second command in the speech lexicon based on hypothesis sequences and error rates of the acoustic engine;
determine a confusability metric corresponding to classification of the first command and the second command by the decoder, the confusability metric indicative of a probability the first command is classified as the second command, the confusability metric based on the renormalized weights;
determine an alternate command corresponding to the first command or the second command, the alternate command phonetically different than the first command or the second command; and
replace the first command or the second command in the speech lexicon with the alternate command responsive to a determination the confusability metric exceeds a threshold value.