US 11,676,585 B1
Hybrid decoding using hardware and software for automatic speech recognition systems
Pushkaraksha Gejji, San Jose, CA (US)
Assigned to Amazon Technologies, Inc., Seattle, WA (US)
Filed by Amazon Technologies, Inc., Seattle, WA (US)
Filed on Aug. 12, 2021, as Appl. No. 17/401,141.
Application 17/401,141 is a continuation of application No. 16/455,400, filed on Jun. 27, 2019, abandoned.
Claims priority of provisional application 62/772,359, filed on Nov. 28, 2018.
Int. Cl. G10L 15/00 (2013.01); G10L 15/18 (2013.01); G10L 15/02 (2006.01); G10L 15/14 (2006.01); G10L 15/19 (2013.01); G10L 15/16 (2006.01); G10L 15/193 (2013.01); G06F 40/289 (2020.01); G10L 15/08 (2006.01)
CPC G10L 15/1822 (2013.01) [G10L 15/02 (2013.01); G10L 15/14 (2013.01); G10L 15/19 (2013.01); G06F 40/289 (2020.01); G10L 15/083 (2013.01); G10L 15/16 (2013.01); G10L 15/18 (2013.01); G10L 15/193 (2013.01)] 18 Claims
OG exemplary drawing
 
1. A method for decoding a speech input, comprising:
receiving speech data from an audio input device based on the speech input;
determining, using a feature extraction module, acoustic features from the speech data; and
determining, using an accelerator and an external processor separate from the accelerator, one or more most-likely words represented in the speech input by:
determining, by the accelerator using a language model, a first word having a first score and a second word having a second score based on the acoustic features;
modifying, by the external processor using a personalized grammar model stored outside of the accelerator, (i) the first score to create a third score and (ii) the second score to create a fourth score;
after creating the third score and the fourth score outside of the accelerator, storing the third score and the fourth score at a local memory of the accelerator; and
determining the one or more most-likely words based on the first word, the third score, the second word, and the fourth score.