Generating confidence scores from word lattices

• US 7,092,883 B1
• Filed: 09/25/2002
• Issued: 08/15/2006
• Est. Priority Date: 03/29/2002
• Status: Expired due to Fees

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: AT&T, Inc.
• Inventors: Gretter, Roberto, Riccardi, Giuseppe
• Primary Examiner(s): Knepper, David D.
• Application Number:
  US10/254,217
• Time in Patent Office: 1,420 Days
• Field of Search: None
• US Class Current: 704/242
• CPC Class Codes: G10L 15/08 Speech classification or search

Abstract

Systems and methods for determining word confidence scores. Speech recognition systems generate a word lattice for speech input. Posterior probabilities of the words in the word lattice are determined using a forward-backward algorithm. Next, time slots are defined for the word lattice, and for all transitions that at least partially overlap a particular time slot, the posterior probabilities of transitions that have the same word label are combined for those transitions. The combined posterior probabilities are used as confidence scores. A local entropy can be computed on the competitor transitions of a particular time slot and also used as a confidence score.

Claims

1. In a system that generates a flot lattice from data input, wherein the flot lattice has one or more transitions that each have a label, a method for determining confidence scores for the labels in the flot lattice, the method comprising:

determining a posterior probability for each transition of the flot lattice;

defining one or more time slots for the slot lattice, wherein one or more transitions is at least partially included in each time slot; and

for each transition that is at least partially included in a particular time slot, combining the posterior probabilities of all transitions that have the same label to obtain a confidence score for that label.

2. A method as defined in claim 1, wherein determining a posterior probability for each transition of the flot lattice further comprises applying a forward-backward algorithm to the flot lattice.

3. A method as defined in claim 1, wherein defining one or more time slots for the slot lattice further comprises defining one or more time slots independently of a topology of the flot lattice.

4. A method as defined in claim 1, wherein each label is a word or a phoneme.

5. A method as defined in claim 1, wherein the one or more time slots correspond to transitions of a best path in the flot lattice.

6. A method as defined in claim 5, wherein determining one or more time slots for the word lattice further comprises identifying a best path from the word lattice, wherein the one or more time slots correspond to the transitions of the best path of the word lattice.

8. A method as defined in claim 1, further comprising identifying ally transitions in each time slot and identifying competitor transitions in each time slot.

9. A method as defined in claim 8, further comprising computing a local entropy on competitor transitions in each time slot.

7. A method as defined in 6, wherein determining one or more time slots for the word lattice further comprises:

applying a forward-backward algorithm to the word lattice; and

combining local likelihoods to determine posterior probabilities for each word label of each transition of the word lattice.

10. In a speech recognition system that generates a word lattice from speech input, wherein the word lattice has one or more transitions, a method for determining word confidence scores for word labels in the word lattice, the method comprising:

applying a forward-backward algorithm to the word lattice to determine a posterior probability for each transition of the word lattice, wherein each transition has a word label;

defining one or more time slots, wherein each time slot has a starting time frame and an ending time frame; and

for transitions of the word lattice that overlap a particular time slot, combining the posterior probabilities of all transitions that have the same word label, wherein the combined posterior probabilities are a confidence score for the word labels of those transitions.

11. A method as defined in claim 10, wherein applying a forward-backward algorithm to the word lattice further comprises combining local likelihoods in order to compute the posterior probabilities of each transition.

12. A method as defined in claim 10, wherein defining one or more time slots further comprises defining one or more time slots using starting time frames and ending time frames that correspond to transitions of a best path of the word lattice.

13. A method as defined in claim 10, wherein defining one or more time slots further comprises defining one or more time slots independently of a topology of the word lattice.

14. A method as defined in claim 10, wherein combining the posterior probabilities of all transitions that have the same word label further comprises identifying both ally transitions and competitor transitions for each transition that overlaps the particular time slot, wherein ally transitions have the same word label.

15. A method as defined in claim 10, wherein combining the posterior probabilities of all transitions that have the same word label further comprises summing posterior probabilities for ally transitions.

16. A method as defined in claim 15, further comprising computing a local entropy on competitor transitions of the particular time slot.

17. In a speech recognition system that does not require training and that generates a word lattice from speech input, wherein the word lattice includes one or more transitions that each have a word label, a method for computing word confidence scores for the word labels that are included in the word lattice, the method comprising:

determining one or more time slots for the word lattice, wherein each transition of the word lattice overlaps at least one time slot and wherein each transition in the word lattice has a posterior probability;

for each time slot;

identifying transitions that have the same word label; and

summing posterior probabilities of transitions that have the same word label to obtain a word confidence score for those word labels.

18. A method as defined in claim 17, wherein identifying transitions that have the same word label further comprises identifying transitions that have the same word label from a set of transitions, wherein the set of transitions includes all transitions that at least partially overlap the time slot.

19. A method as defined in claim 17, wherein summing posterior probabilities of transitions further comprises summing posterior probabilities of ally transitions.

20. A method as defined in claim 17, further comprising computing a local entropy on competitor transitions of a particular time slot, wherein competitor transitions have different word labels.

21. A method as defined in claim 17, wherein determining one or more time slots for the word lattice further comprises determining one or more time slots independently of a topology of the word lattice.

22. A computer program product having computer-executable instructions for performing elements of claim 17.

23. In a speech recognition system that generates a word lattice from speech input, wherein the word lattice has one or more transitions that each have a word label, a method for determining word confidence scores for word labels in the word lattice, the method comprising:

defining one or more time slots that correspond to transitions of a best path of the word lattice, wherein each time slot has a starting time frame and an ending time frame;

identifying ally transitions and competitor transitions for each time slot, wherein ally transitions have the same word label and wherein competitor transitions have different labels; and

computing a local entropy on competitor transitions in each time slot, wherein the local entropy is a confidence score for the word labels of the competitor transitions.

24. A method as defined in claim 23, wherein applying a forward-backward algorithm to the word lattice further comprises combining local likelihoods in order to compute the posterior probabilities of each transition.

25. A method as defined in claim 23, wherein defining one or more time slots that correspond to transitions of a best path of the word lattice further comprises applying a forward-backward algorithm to the word lattice to determine a posterior probability for each transition of the word lattice, wherein each transition has a word label.

26. A method as defined in claim 25, wherein identifying ally transitions and competitor transitions for each time slot further comprises identifying ally transitions and competitor transitions for each time slot from all transitions that at least partially overlap each time slot.

27. A method as defined in claim 26, further comprising combining posterior probabilities of ally transitions within a time slot, wherein the combined posterior probabilities are confidence scores for those word labels.

28. A method as defined in claim 23, wherein defining one or more time slots further comprises defining one or more time slots independently of a topology of the word lattice.

29. A computer program product having computer-executable instructions for performing the elements of claim 23.

30. In a speech recognition system that generates a word lattice from speech input, wherein the word lattice has one or more transitions that each have a word label, a computer program product for implementing a method for determining word confidence scores for word labels in the word lattice, the method comprising:

a computer-readable medium having computer-executable instructions for performing the method, the method comprising;

applying a forward-backward algorithm to the word lattice to determine a posterior probability for each transition of the word lattice, wherein each transition has a word label;

defining one or more time slots, wherein each time slot has a starting time frame and an ending time frame; and

for transitions of the word lattice that overlap a particular time slot, combining the posterior probabilities of all transitions that have the same word label, wherein the combined posterior probabilities are a confidence score for those word labels.

31. A computer program product as defined in claim 30, wherein applying a forward-backward algorithm to the word lattice further comprises combining local likelihoods in order to compute the posterior probabilities of each transition.

32. A computer program product as defined in claim 30, wherein defining one or more time slots further comprises defining one or more time slots using starting time frames and ending time frames that correspond to transitions of a best path of the word lattice.

33. A computer program product as defined in claim 30, wherein defining one or more time slots further comprises defining one or more time slots independently of a topology of the word lattice.

34. A computer program product as defined in claim 30, wherein combining the posterior probabilities of all transitions that have the same word label further comprises identifying both ally transitions and competitor transitions for each transition that overlaps the particular time slot, wherein ally transitions have the same word label.

35. A computer program product as defined in claim 30 further comprising computing a local entropy on transitions of the particular time slot that do not have the same word label.