Distributed pattern recognition training method and system

• US 20060015341A1
• Filed: 07/14/2005
• Published: 01/19/2006
• Est. Priority Date: 07/15/2004
• Status: Active Grant

• Current Assignee: Aurilab LLC
• Original Assignee: Aurilab LLC
• Inventors: Baker, James K.
• Granted Patent: US 7,562,015 B2
• Time in Patent Office: Days
• US Class Current: 704/255
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/30 Distributed recognition, e.g. in client-server systems, for mobile phones or network applications

Abstract

A distributed pattern recognition training method includes providing data communication between at least one central pattern analysis node and a plurality of peripheral data analysis sites. The method also includes communicating from the at least one central pattern analysis node to the plurality of peripheral data analysis a plurality of kernel-based pattern elements. The method further includes performing a plurality of iterations of pattern template training at each of the plurality of peripheral data analysis sites.

Claims

1. A distributed pattern recognition training method, comprising:

providing data communication between at least one central pattern analysis node and a plurality of peripheral data analysis sites;

communicating, from said at least one central pattern analysis node to said plurality of peripheral data analysis sites, a plurality of kernel-based pattern elements; and

performing a plurality of iterations of pattern template training at each of said plurality of peripheral data analysis sites.

2. The distributed pattern recognition training method according to claim 1, wherein each iteration of pattern template training comprises:

communicating, from said at least one central pattern analysis node, a candidate solution to a pattern discrimination problem computable from said plurality of kernel-based pattern elements;

at each of said plurality of peripheral data analysis sites, obtaining a plurality of data items; and

at each of said plurality of peripheral data analysis sites, computing statistics derived from said candidate solution, said plurality of kernel-based pattern elements and said plurality of data items.

3. The distributed pattern recognition training method according to claim 2, wherein each iteration of pattern template training further comprises:

communicating said statistics from said plurality of peripheral data analysis sites to said central pattern analysis node; and

performing a pattern template analysis at said central pattern analysis node to update said candidate solution based at least in part on said statistics communicated from said peripheral data analysis sites.

4. The distributed pattern recognition training method according to claim 3, further comprising:

for at least one of said iterations of pattern template training, communicating at least one candidate kernel-based pattern element from at least one of said peripheral data analysis sites to said central pattern analysis node; and

for at least one of said iterations of pattern template training, performing a selection operation among said plurality of kernel-based pattern elements and said at least one candidate kernel-based pattern element to change the number of said kernel-based pattern elements used in said candidate solution to be communicated to said plurality of peripheral data analysis sites in the next iteration of pattern template training.

5. The distributed pattern recognition training method according to claim 1, wherein said pattern recognition is speech recognition.

6. The distributed pattern recognition training method according to claim 1, wherein said pattern recognition is classification of the output of a speech recognition system.

7. The distributed pattern recognition training method according to claim 1, further comprising:

obtaining a set of data observations at at least one of said plurality of peripheral data analysis sites, wherein said set of data observations include acoustic feature observations.

8. The distributed pattern recognition training method according to claim 3, further comprising:

obtaining a set of data observations at at least one of said plurality of peripheral data analysis sites, wherein said set of data observations include sequences of words.

9. The distributed pattern recognition training method according to claim 1, wherein said plurality of kernel-based pattern elements communicated from said central pattern analysis node to said plurality of peripheral data analysis sites includes a plurality of data items.

10. The distributed pattern recognition training method according to claim 9, further comprising creating a plurality of functionals based on said plurality of data items and at least one reproducing kernel function.

11. The distributed pattern recognition training method according to claim 1, wherein said plurality of kernel-based pattern elements communicated from said central pattern analysis node to said plurality of peripheral data analysis sites includes a plurality of functionals.

12. The distributed pattern recognition training method according to claim 1, further comprising communicating a plurality of reproducing kernel functions from said central pattern analysis node to said plurality of peripheral data analysis sites.

13. The distributed pattern recognition training method according to claim 3, wherein said pattern template analysis performed at said central pattern analysis node includes solving a constrained optimization problem.

14. The distributed pattern recognition training method according to claim 13, wherein said constrained optimization problem is a convex programming problem.

15. The distributed pattern recognition training method according to claim 13, wherein said constrained optimization problem is a linear programming problem.

16. The distributed pattern recognition training method according to claim 13, wherein said constrained optimization problem is a quadratic programming problem.

17. The distributed pattern recognition training method according to claim 1, wherein said plurality of kernel-based pattern elements contains at least one data item,

the method further comprising:

matching said plurality of kernel-based pattern elements to a plurality of peripheral data items.

18. The distributed pattern recognition training method according to claim 17, wherein the matching step comprises computing the distance between said at least one data item from said plurality of kernel-based pattern elements and each of a plurality of peripheral data items based on some distance metric.

19. The distributed pattern recognition training method according to claim 18, wherein said distance metric is the Euclidean distance.

20. The distributed pattern recognition training method according to claim 18, wherein said distance metric is a reproducing kernel in a Hilbert space.

45. The distributed pattern recognition training method according to claim 1, wherein at least one of said plurality of kernel-based pattern elements is a data item which is equal to a selected example data item in a set of training data.

46. The distributed pattern recognition training method according to claim 1, wherein at least one of said plurality of kernel-based pattern elements is the mean vector of a plurality of data items.

47. The distributed pattern recognition training method according to claim 10, wherein at least one of said plurality of reproducing kernel functions is a function of the covariance matrix of a plurality of data vectors.

48. The distributed pattern recognition training method according to claim 3, wherein said statistics communicated from a particular one of said plurality of peripheral data analysis sites include a vector proportional to the direction in which to change one of said kernel-based pattern elements to get the greatest rate of improvement of a specified objective function.

49. The distributed pattern recognition training method according to claim 10, wherein said statistics communicated from a particular one of said plurality of peripheral data analysis sites include a vector proportional to the direction in which to change at least one of said reproducing kernel functions to get the greatest rate of improvement of the degree of fit between said plurality of peripheral data items obtained at said particular one of said data analysis sites and a family of functionals based on said at least one reproducing kernel function.

50. The distributed pattern recognition training method according to claim 3, wherein said performing a pattern template analysis at said central analysis node further comprises solving a constrained optimization problem for a selected subset of said kernel-based pattern elements.

51. The distributed pattern recognition training method according to claim 1, further comprising computing constraint satisfaction statistics for kernel-based pattern elements obtained at at least one of said peripheral data analysis sites.

52. The distributed pattern recognition training method according to claim 1, further comprising providing said data communication between said central pattern analysis node and said plurality of peripheral data analysis sites at least in part by utilizing a computer network.

53. The distributed pattern recognition training method according to claim 1, further comprising providing said data communication between said central pattern analysis node and said plurality of peripheral data analysis sites at least in part by physically transferring data storage media between said central pattern analysis node and said plurality of peripheral data analysis sites.

21. A speech recognition method comprising:

performing a base speech recognition process which generates utterance hypotheses;

obtaining a representation of a set of event types which may occur in utterance hypotheses generated by said base speech recognition process;

obtaining a plurality of hypothesized event pair discrimination models;

obtaining a plurality of utterance hypotheses from said base speech recognition process;

selecting at least one pair from said plurality of utterance hypotheses;

for each selected pair of utterance hypotheses, selecting at least one point in time such that within a specified interval of said point in time a first hypothesized particular event happens according to a first hypothesis of said pair of utterance hypotheses and a second hypothesized particular event happens according to a second hypothesis of said pair of utterance hypotheses;

obtaining data observations at said at least one point in time;

rescoring said pair of utterance hypotheses based at least in part on said event pair discrimination models and said data observations at said at least one point in time; and

re-ranking said plurality of hypotheses based on said rescoring of said selected at least one pair from said plurality of hypotheses.

22. The speech recognition method according to claim 21, wherein each discrimination model discriminates between one particular event type that may happen according to a first candidate hypothesis generated by said base recognition process and a different particular event type that may happen within a specified interval of the same point in time according to a second candidate hypothesis generated by said base recognition process.

23. The speech recognition method according to claim 21, wherein in said plurality of discrimination models there is a model which discriminates between said first hypothesized particular event and said second hypothesized particular event.

24. The speech recognition method according to claim 21, wherein at least one of said discrimination models is obtained by training a classification system on samples of errors made by said base speech recognition process.

25. A speech recognition system according to claim 24, wherein said training of said classification system comprises distributed pattern recognition training which includes:

providing data communication between at least one central pattern analysis node and a plurality of peripheral data analysis sites;

communicating at least one kernel-based pattern element from said central pattern analysis node to said plurality of peripheral data analysis sites;

at each of said plurality of peripheral data analysis sites obtaining acoustic observation data;

communicating statistics from said plurality of peripheral data analysis sites to said central pattern analysis node; and

updating said at least one kernel-based pattern element based on said statistics.

26. The speech recognition system according to claim 21, wherein said set of event types includes at least one event of the form that a particular phoneme occurs during a time interval that overlaps with a given point in time.

27. The speech recognition system according to claim 21, wherein said set of event types includes at least one event of the form that an acoustic segment boundary of a specified type occurs within a specified interval of a given point in time.

28. The speech recognition system according to claim 21, wherein said set of event types includes at least one event of the form that a particular word occurs at a particular point in a word sequence.

29. The speech recognition system according to claim 28, further comprising representing each word as a vector of features.

30. The speech recognition system according to claim 28, further comprising using the respective context of hypothesized words from each of said one hypothesis and said second hypothesis.

31. A two stage speech recognition method, comprising:

obtaining a base recognition process which generates utterance hypotheses;

obtaining a representation of the set of event types which might occur in utterance hypotheses generated by said base speech recognition process; and

obtaining a plurality of self-normalizing event detection verification models trained at least in part on errors made by said base speech recognition system;

obtaining a plurality of hypotheses from said base speech recognition system;

for each of said plurality of hypotheses, obtaining a list of hypothesized events which happen according to said hypothesis and the hypothesized time at which each of said events occurs;

rescoring each of said plurality of hypotheses by adding the output score from the event detection verification model for each event in said list of hypothesized events; and

re-ranking said plurality of hypotheses based on said rescoring and basing the output of said two stage speech recognition method on said re-ranking.

32. The two stage speech recognition method according to claim 31,

wherein said event detection verification model for a particular event is trained to attempt to produce an output score of a particular numeric sign whenever said base recognition system incorrectly scores a hypothesis not containing said particular event as better than a hypothesis containing said particular event, and

wherein said event detection verification model for said particular event is trained to attempt to produce an output score of the opposite numeric sign whenever said base recognition system incorrectly scores a hypothesis containing said particular event as better than a hypothesis not containing said particular event;

33. The two stage speech recognition method according to claim 31, wherein the number of elements in said list of hypothesized events for a particular hypothesis depends only on said particular hypothesis and is different from the number of elements in said list of hypothesized events for a different hypothesis.

34. The two stage speech recognition method according to claim 33, wherein said set of event types include at least one event of the form of a word occurring in the particular context of other words in a particular hypothesis.

35. A two stage speech recognition method according to claim 33, wherein said set of event types include at least one event of the form of a particular kind of acoustic segment boundary.

36. A pattern scoring method, comprising:

obtaining a plurality of template data items;

obtaining a plurality of kernel functions;

for each of said plurality of template data items, creating a plurality of functionals where each particular functional is associated with a particular template data item and a particular kernel function;

computing the score for each sample data item based on the value of a linear combination of a subset of said plurality of functionals; and

for each pattern to be scored, selecting the particular functionals to be used and the weight to be given to each particular functional based on a constrained optimization problem which minimizes a function of the weights for a given amount of separation between the pattern classes.

37. The pattern scoring method according to claim 36, wherein said kernel functions are reproducing kernels in a Hilbert space.

38. The pattern scoring method according to claim 36, wherein said data items are word sequences.

39. The pattern scoring method according to claim 38, wherein said obtaining of a plurality of kernel functions includes obtaining a plurality of subsets of word context positions and obtaining a metric based on word comparisons in each of the plurality of the subsets of word context positions.

40. The pattern scoring method according to claim 39, wherein at least one of said subsets of word context positions includes at least one position earlier in the word sequence than a designated word position and at least one of said subsets of word context positions includes at least one position later in the word sequence than said designated word position.

41. The pattern scoring method according to claim 38, wherein said plurality of kernel functions is based at least in part by parsing said word sequences according to a specified grammar.

42. The pattern scoring method according to claim 38, wherein said plurality of kernel functions is based at least in part on the estimated state sequence of a hidden stochastic process generating each given observed word sequence.

43. The pattern scoring method according to claim 36, wherein said constrained optimization problem is a linear programming problem.

44. The pattern scoring method according to claim 36, wherein said constrained optimization problem is a quadratic programming problem.

54. A program product having machine-readable program code for performing distributed pattern recognition training, the program code, when executed, causing a machine to perform the following steps:

providing data communication between at least one central pattern analysis node and a plurality of peripheral data analysis sites;

communicating, from said at least one central pattern analysis node to said plurality of peripheral data analysis sites, a plurality of kernel-based pattern elements; and

performing a plurality of iterations of pattern template training at each of said plurality of peripheral data analysis sites.

55. The program product according to claim 54, wherein each iteration of pattern template training comprises:

communicating, from said at least one central pattern analysis node, a candidate solution to a pattern discrimination problem computable from said plurality of kernel-based pattern elements;

at each of said plurality of peripheral data analysis sites, obtaining a plurality of data items; and

at each of said plurality of peripheral data analysis sites, computing statistics derived from said candidate solution, said plurality of kernel-based pattern elements and said plurality of data items.

56. The program product according to claim 55, wherein each iteration of pattern template training further comprises:

communicating said statistics from said plurality of peripheral data analysis sites to said central pattern analysis node; and

performing a pattern template analysis at said central pattern analysis node to update said candidate solution based at least in part on said statistics communicated from said peripheral data analysis sites.

57. The program product according to claim 55, further comprising:

for at least one of said iterations of pattern template training, communicating at least one candidate kernel-based pattern element from at least one of said peripheral data analysis sites to said central pattern analysis node; and

for at least one of said iterations of pattern template training, performing a selection operation among said plurality of kernel-based pattern elements and said at least one candidate kernel-based pattern element to change the number of said kernel-based pattern elements used in said candidate solution to be communicated to said plurality of peripheral data analysis sites in the next iteration of pattern template training.

58. The program product according to claim 54, wherein said pattern recognition is speech recognition.

59. The program product according to claim 54, wherein said pattern recognition is classification of the output of a speech recognition system.

60. The program product according to claim 54, further comprising:

obtaining a set of data observations at at least one of said plurality of peripheral data analysis sites, wherein said set of data observations include acoustic feature observations.