US 12,254,866 B2
Systems and methods for aligning a reference sequence of symbols with hypothesis requiring reduced processing and memory
Jean-Philippe Robichaud, Mercier (CA); Miguel Jette, Squamish (CA); Joshua Ian Dong, Austin, TX (US); Quinten McNamara, Austin, TX (US); Nishchal Bhandari, Austin, TX (US); and Michelle Kai Yu Huang, Austin, TX (US)
Assigned to Rev.com, Inc., San Francisco, CA (US)
Filed by Rev.com, Inc., San Francisco, CA (US)
Filed on Oct. 13, 2020, as Appl. No. 17/069,462.
Prior Publication US 2022/0115003 A1, Apr. 14, 2022
Int. Cl. G10L 15/06 (2013.01); G10L 15/14 (2006.01)
CPC G10L 15/06 (2013.01) [G10L 15/14 (2013.01)] 1 Claim
OG exemplary drawing
 
1. A system for determining an alignment sequence between a reference sequence of symbols and a hypothesis sequence of symbols, the system comprising:
a loading module executing code and configured to:
load a reference sequence of symbols to a computing system;
load a hypothesis sequence of symbols to the computing system;
load a reference sequence of symbols meta-data;
load hypothesis sequence of symbols meta-data;
a finite state automaton module executing code and configured to:
create a reference finite state automaton for the reference sequence of symbols;
create a hypothesis finite state automaton for the hypothesis sequence of symbols;
a traversing module executing code and configured to:
traverse the reference finite state automaton, adding new reference arcs and new reference transforming properties;
create arcs in the reference finite state automaton for the reference sequence of symbols meta-data;
create arcs in the hypothesis finite state automaton for the hypothesis sequence of symbols meta-data, wherein the hypothesis sequence of symbols meta-data is representative of at least one property of the hypothesis sequence of symbols and each arc of the hypothesis finite state automaton represents a word;
traverse the hypothesis finite state automaton, adding new hypothesis arcs and new hypothesis transforming properties;
a composing module, executing code and configured to:
compose the hypothesis finite state automaton with the reference finite state automaton creating alternative paths to form a composed finite state automaton;
track a number of the alternative paths created;
prune the alternative paths based on likely top paths;
a scoring module, executing code and configured to:
backtrack over most likely paths of the composed finite state automaton;
rescore edit-distances of the composed finite state automaton;
determine an alignment sequence of the reference sequence of symbols and the reference sequence of symbols meta-data based the edit-distances of the composed finite state automaton; wherein the reference sequence of symbols meta-data is representative of at least one property of the reference sequence of symbols, the reference sequence of symbols is representative of one of transcribed speech and translated speech, and each arc of the reference finite state automaton represents a word.