N-best list rescoring in speech recognition
View Patent ↗A method of speech recognition processing is described based on an N-best list of recognition hypotheses corresponding to a spoken input. Each hypothesis on the N-best list is rescored based on its rank in the rescored N-best list. The rescoring may be based on a Statistical Language Model (SLM) or Dynamic Semantic Model (DSM). One or more rescoring categories may be associated with each recognition hypotheses to affect or bias the rescoring.
1. A method of speech recognition processing comprising:
in an automatic speech recognition engine:
producing with a pattern matching recognizer an N-best list of recognition hypotheses corresponding to a spoken input; and
rescoring the hypotheses in the automatic speech recognition engine to produce a rescored N-best list output from the automatic speech recognition engine;
wherein the rescoring uses a plurality of rescoring categories based on position in the rescored N-best list such that some positions in the rescored N-best list are rescored based on a first combination of rescoring categories and other positions in the rescored N-best list are rescored based on a second combination of rescoring categories.
2. A method according to claim 1 , wherein the rescoring is based on using a Statistical Language Model (SLM).
3. A method according to claim 1 , wherein the rescoring is based on using a Dynamic Semantic Model (DSM).
4. A method according to claim 1 , wherein the rescoring includes applying a bias to each rescored hypothesis depending on its rank in the rescored N-best list.
5. A method according to claim 1 , wherein the at least one rescoring category includes a category for most recently used recognition hypotheses.
6. A method according to claim 1 , wherein the at least one rescoring category includes a category for most frequently used recognition hypotheses.
7. A method according to claim 6 , wherein the at least one rescoring category includes a category for names within a geographic vicinity of one or more most frequently used names.
8. A method according to claim 1 , in which selected positions in the N-best list are reserved for recognition hypotheses of one or more selected rescoring categories.
9. A method according to claim 8 , wherein producing an N-best list initially considers only hypotheses in the one or more selected rescoring categories.
10. A method according to claim 1 , further comprising:
providing a first output of the rescored hypotheses in a selected number of the top positions in the rescored N-best list; and
in response to a user action, providing a second output of the remaining rescored hypotheses.
11. A method according to claim 1 , wherein the recognition hypotheses represent place names for a navigation system.
12. A method according to claim 11 , wherein the place names are city names.
13. A method according to claim 11 , wherein the place names are street names.
14. A method according to claim 1 , wherein the rescoring includes:
dividing the rescored N-best list into blocks, each block corresponding to a range of ranks in the rescored N-best list, the block boundaries varying depending on a metric corresponding to an expected recognition accuracy for the spoken input.
15. A method according to claim 14 , wherein the metric is based on a signal-to-noise ratio.
16. A speech recognition processing arrangement comprising:
an automatic speech recognition engine having:
means for producing an N-best list of recognition hypotheses corresponding to a spoken input; and
means for rescoring the hypotheses to produce a rescored N-best list output;
wherein the rescoring uses a plurality of rescoring categories based on position in the rescored N-best list such that some positions in the rescored N-best list are rescored based on a first combination of rescoring categories and other positions in the rescored N-best list are rescored based on a second combination of rescoring categories.
17. A speech processing arrangement according to claim 16 , wherein the means for rescoring is based on using Statistical Language Model (SLM) means.
18. A speech processing arrangement according to claim 16 , wherein the means for rescoring is based on using Dynamic Semantic Model (DSM) means.
19. A speech processing arrangement according to claim 16 , wherein the means for rescoring includes applying a bias to each rescored hypothesis depending on its rank in the rescored N-best list.
20. A speech processing arrangement according to claim 16 , wherein the at least one rescoring category includes a category for most recently used recognition hypotheses.
21. A speech processing arrangement according to claim 16 , wherein the at least one rescoring category includes a category for most frequently used recognition hypotheses.
22. A speech processing arrangement according to claim 21 , wherein the at least one rescoring category includes a category for names within a geographic vicinity of one or more most frequently used names.
23. A speech processing arrangement according to claim 16 , in which selected positions in the N-best list are reserved for recognition hypotheses of one or more selected rescoring categories.
24. A speech processing arrangement according to claim 23 , wherein the means for producing an N-best list initially considers only hypotheses in the one or more selected rescoring categories.
25. A speech processing arrangement according to claim 16 , further comprising:
means for providing a first output of the rescored hypotheses in a selected number of the top positions in the rescored N-best list; and
means for providing, in response to a user action, a second output of the remaining rescored hypotheses.
26. A speech processing arrangement according to claim 16 , wherein the recognition hypotheses represent place names for a navigation system.
27. A speech processing arrangement according to claim 26 , wherein the place names are city names.
28. A speech processing arrangement according to claim 26 , wherein the place names are street names.
29. A speech processing arrangement according to claim 16 , wherein the means for rescoring includes means for dividing the rescored N-best list into blocks, each block corresponding to a range of ranks in the rescored N-best list, the block boundaries varying depending on a metric corresponding to an expected recognition accuracy for the spoken input.
30. A speech processing arrangement according to claim 29 , wherein the metric is based on a signal-to-noise ratio.