Pattern matching method, apparatus and computer readable memory medium for speech recognition using dynamic programming

• US 5,960,395 A
• Filed: 02/06/1997
• Issued: 09/28/1999
• Est. Priority Date: 02/09/1996
• Status: Expired due to Term

• Current Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Original Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Inventors: Tzirkel-Hancock, Eli
• Primary Examiner(s): Dorvil, Richemond
• Application Number:
  US08/796,420
• Time in Patent Office: 964 Days
• Field of Search: 704/241, 704/231, 704/236, 704/243, 704/244, 704/245, 704/238, 704/255, 704/256, 704/240, 704/257
• US Class Current: 704/241
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]
• CPC Class Codes: G10L 2015/085 Methods for reducing search complexity, pruning
• CPC Class Codes: H04M 1/271 controlled by voice recognition
• CPC Class Codes: H04M 1/57 Arrangements for indicating or recording the number of the calling subscriber at the called subscriber's set at the operator set in a manual exchange H04M5/20
• CPC Class Codes: H04M 1/6505 storing speech in digital form
• CPC Class Codes: H04M 2201/40 using speech recognition
• CPC Class Codes: H04M 3/42204 Arrangements at the exchange for service or number selection by voice at the terminal H04M1/27
• CPC Class Codes: H04M 3/533 Voice mail systems
• CPC Class Codes: H04M 3/53383 Message registering commands or announcements; Greetings

Abstract

A method for matching an input pattern with a number of stored reference patterns using a dynamic programming matching technique is described. The reference patterns of a reference signal which are at the end of a dynamic programming path for a current input pattern are listed in an active list. The dynamic programming paths are propagated by processing the reference patterns on the active list, and a new active list is generated for the succeeding input pattern. The amount of processing required for each pattern on the active list is reduced by using a pointer which identifies the reference pattern which is the earliest in the sequence of patterns of the current reference signal listed on the new active list during the processing of a preceding dynamic programming path. In a second aspect, a speech recognition interface is used as a control system for a telephony system.

Claims

1. An apparatus for performing a dynamic programming pattern matching process, the apparatus comprising:

input means for inputting a sequence of input patterns representative of an input signal;

memory means for storing a plurality of sequences of references patterns, each sequence being representative of a reference signal;

means for processing a current input pattern with respect to at least some of the reference signals in turn, comprising;

(1) means for defining as active patterns the reference patterns of a current reference signal which are at the end of a dynamic programming path for the current into pattern being processed, each path representing a possible matching between an ordered sequence of reference patterns and an ordered sequence of input patterns ending at said current input pattern, and for listing the active patterns for the current input pattern in a current active list;

(2) means for storing in a store associated with each active pattern, a cumulative value representative of a score for the dynamic programming path which ends at that active pattern for said current input pattern; and

(3) means for processing each active pattern of said current reference signal in reverse sequential order, to update said cumulative values and to propagate said dynamic programming paths based on constraints which are placed on the dynamic programming path propagation, said active pattern processing means comprising;

(A) means for updating the cumulative value stored in the store associated with a current active pattern being processed, using said current input pattern; and

(B) means for propagating the dynamic programming path associated with the current active pattern, and for listing, if it is not already listed, each reference pattern of the current reference signal, which may be at the end of that dynamic programming path for the succeeding input pattern, in a new active list; and

wherein said propagating means is controlled using a pointer associated with the current reference signal, which identifies the reference pattern which is the earliest in the sequence of patterns of the current reference signal listed in the new active list, after the processing of the preceding active pattern, such that the propagating means can propagate each dynamic programming path without the need to search the new active list to identify which reference patterns, of the current reference signal, have been listed on the new active list as a result of said active pattern processing means processing preceding active patterns of the current reference signal.

2. An apparatus according to claim 1, wherein the propagating means comprises:

(i) means for determining which of the reference patterns of the current reference signal are potentially at an end of the dynamic programming path which ends at said current active pattern for said succeeding input pattern, based upon said constraints which are placed on the dynamic programming path propagation; and

(ii) means for copying appropriate cumulative values into stores associated with the reference patterns determined by said determining means.

3. An apparatus according to claim 2, wherein said propagating means further comprises means for processing each of the reference patterns determined by said determining means comprising:

means for discriminating whether a current reference pattern being processed is sequentially beyond any reference pattern identified by said pointer; and

means, responsive to said discriminating means, for (i) copying the cumulative value stored in the store associated with said current active pattern into the store associated with the current reference pattern, if it is better than the cumulative value already stored there and if the current reference pattern is the same as or sequentially beyond the identified reference pattern; and

(ii) copying the cumulative value stored in the store associated with said current active pattern into the store associated with said current reference pattern, and listing said current reference pattern in the new active list, if the current reference pattern is sequentially before the identified reference pattern or if there are no reference patterns of the current reference signal in the new active list.

4. An apparatus according to claim 3, wherein said reference pattern processing means is operable to process said reference patterns determined by said determining means in reverse sequential order.

5. An apparatus according to claim 2, wherein said propagating means is operable to list the current active pattern in the new active list, if this is not contrary to any constraint placed on the dynamic programming process if one of the reference patterns determined by said determining means is the current active pattern.

6. An apparatus according to claim 1, further comprising means for setting the pointer associated with the current reference signal to identify the reference pattern on the new active list which is the earliest in the sequence of patterns of the current reference signal once the new active list has been updated by the processing of the current active pattern by said active pattern processing means.

7. An apparatus according to claim 1, wherein the input pattern processing means is operable to process the current input pattern before the succeeding input pattern is received.

8. An apparatus according to claim 1, wherein said input means is operable to input the whole input signal before the first input pattern is processed by said input pattern processing means.

9. An apparatus according to claim 1, further comprising a current active list and a new active list for each reference signal.

10. An apparatus according to claim 1, wherein said propagating means is operable to list said reference patterns in said new active list in reverse sequential order.

11. An apparatus according to claim 1, further comprising means for preventing said active pattern processing means from processing the current active pattern, if the cumulative value is associated therewith is worse than a preset threshold value.

12. An apparatus according to claim 11, further comprising means for varying said threshold value in order to try to maintain the number of active patterns processed by said active pattern processing means for each input pattern processed by said input pattern processing means to be below a given maximum number.

13. An apparatus according to claim 12, wherein the threshold value varying means is operable to set the threshold value used during the processing of the succeeding input pattern during the processing of the current input pattern, and is dependent upon the total number of reference patterns in the new active list or lists for the succeeding input pattern.

14. An apparatus according to claim 1, further comprising constraining means for constraining said dynamic programming path propagation so that a limit is placed on the number of consecutive input patterns that a reference pattern can be matched with, on the same dynamic programming path.

15. An apparatus according to claim 1, further comprising constraining means for constraining said dynamic programming path propagation so that it a reference pattern is matched with an input pattern, then a limited number of reference patterns which are sequentially beyond that reference pattern can be matched with the succeeding input pattern, on the same dynamic programming path.

16. An apparatus according to claim 1, further comprising constraining means for constraining said dynamic programming path propagation so that if a reference pattern is matched with an input pattern, then reference patterns which are sequentially before that reference pattern can not be matched with the succeeding input pattern, on the same dynamic programming path.

17. An apparatus according to claim 1, wherein said reference patterns are representative of a template, and wherein said cumulative values are distance measures.

18. An apparatus according to claim 1, wherein said reference patterns are representative of a statistical model, and wherein said cumulative values are probabilistic measures.

19. An apparatus according to claim 1, wherein said apparatus is arranged so that said input signal can be matched with a sequence of said reference signals, by allowing dynamic programming paths propagating in one reference signal to subsequently propagate into other reference signals.

20. An apparatus according to claim 19, wherein said apparatus is arranged so that only the dynamic programming path, which has the best cumulative value of all the dynamic programming paths which leave the reference pattern for the current input pattern, propagates further.

21. An apparatus according to claim 19, further comprising means for using a set of defined rules for constraining the sequence of reference signals against which the input signal can be matched.

22. An apparatus according to claim 19, further comprising reference signal processing means operable, after the current input pattern has been processed with respect to the reference signals by said input pattern processing means, to process those reference signals to update the new active list and to update the cumulative values associated with the reference patterns at the beginning of those reference signals, which are potentially at an end of a dynamic programming path which enters the corresponding reference signal, based upon said constraints, using the best cumulative value of all the dynamic programming paths which can enter that reference signal.

23. An apparatus according to claim 1, wherein said input signal is a speech signal, and wherein each input pattern comprises a number of parameters representative of the acoustic properties of the speech signal during a corresponding time frame.

24. An apparatus according to claim 23, wherein said input speech signal can be matched to a sequence of reference speech signals, and wherein the allowed sequences are defined by a language model.

25. An apparatus according to claim 24, wherein said apparatus is arranged so that the end of a dynamic programming path which leaves a reference signal when processing the current input pattern can be matched with a pattern representative of background noise for succeeding input patterns.

117. A method according to claim 107, wherein the method only performs step (3) of claim 1 for the current active pattern, if the cumulative value associated therewith is better than a threshold value.

118. A method according to claim 117, wherein said threshold value is varied to try to maintain the number of active patterns processed for each input pattern to be below a given maximum number.

119. A method according to claim 118, wherein the threshold value used during the processing of the succeeding input pattern is determined during the processing of the current input pattern, and is dependent upon the total number of reference patterns in the new active list or lists for the succeeding input pattern.

26. A speech recognition apparatus for recognizing an input speech signal by comparing it with a plurality of reference speech signals, the apparatus comprising:

means for extracting a sequence of input patterns representative of an input speech signal;

means for storing sequences of reference patterns, each sequence being representative of a corresponding reference speech signal;

means for matching the sequence of input patterns with the reference speech signal, using an apparatus for performing a dynamic programming pattern matching process, the dynamic-programming-pattern-matching-process apparatus comprising;

input means for inputting a sequence of input patterns representative of an input signal;

memory means for storing a plurality of sequences of references patterns, each sequence being representative of a reference signal; and

means for processing a current input pattern with respect to at least some of the reference signals in turn, comprising;

(1) means for defining as active patterns the reference patterns of a current reference signal which are at the end of a dynamic programming path for the current input pattern being processed, each path representing a possible matching between an ordered sequence of reference patterns and an ordered sequence of input patterns ending at said current input pattern, and for listing the active patterns for the current input pattern in a current active list;

(2) means for storing in a store associated with each active pattern, a cumulative value representative of a score for the dynamic programming path which ends at that active pattern for said current input pattern; and

(3) means for processing each active pattern of said current reference signal in reverse sequential order, to update said cumulative values and to propagate said dynamic programming paths based on constraints which are placed on the dynamic programming path propagation, said active pattern processing means comprising;

(a) means for updating the cumulative value stored in the store associated with a current active pattern being processed, using said current input pattern; and

(b) means for propagating the dynamic programming path associated with the current active pattern, and for listing, if it is not already listed, each reference pattern of the current reference signal, which may be at the end of that dynamic programming path for the succeeding input pattern, in a new active list; and

wherein said propagating means is controlled using a pointer associated with the current reference signal, which identifies the reference pattern which is the earliest in the sequence of patterns of the current reference signal listed in the new active list, after the processing of the preceding active pattern, such that the propagating means can propagate each dynamic programming path without the need to search the new active list to identify which reference patterns, of the current reference signal, have been listed on the new active list as a result of said active pattern processing means processing preceding active patterns of the current reference signal, and wherein said input signal is a speech signal, and wherein each input pattern comprises a number of parameters representative of the acoustic properties of the speech signal during a corresponding time frame; and

means for providing a recognition result from cumulative values determined by said matching means.

27. An apparatus according to claim 26, wherein said providing means is operable to provide said recognition result by determining which of the dynamic programming paths which end at the last input pattern in the sequence has the best cumulative value.

28. A computer-readable memory medium storing a computer-readable program embodied therein for instructing a computer to perform a dynamic programming pattern matching process between a sequence of input patterns representative of an input signal, and a number of stored sequences of reference patterns, each sequence being representative of a reference signal, wherein the program instructs the computer to process each input pattern in turn with respect to at least some of the reference signals, said program comprising:

(1) a defining instructing means for instructing the computer to define as active patterns the reference patterns of a current reference signal which are at the end of a dynamic programming path for a current input pattern being processed, each path representing a possible matching between an ordered sequence of reference patterns and an ordered sequence of input patterns ending at said current input pattern, and for instructing the computer to list the active patterns for the current input pattern in a current active list;

(2) a storing instructing means for instructing the computer to, for each active pattern, store in a store associated with that active pattern, a cumulative value representative of a score for the dynamic programming path which ends at that active pattern for said current input pattern; and

(3) an updating and propagating instructing means for instructing the computer to update said cumulative values and to propagate said dynamic programming paths based on constraints which are placed on the dynamic programming path propagation, by processing each active pattern of said current reference signal in reverse sequential order, comprising;

(A) an updating instructing means for instructing the computer to update the cumulative value stored in the store associated with a current active pattern being processed, using said current input pattern; and

(b) (B) a first propagating instructing means for instructing the computer to propagate the dynamic programming path associated with the current active pattern, and to list, if it is not already listed, each reference pattern of the current reference signal, which may be at the end of that dynamic programming path for the succeeding input pattern, in a new active list; and

wherein the propagating operation for each dynamic programming path for a current reference signal is controlled using a pointer associated with the current reference signal, which identifies the reference pattern which is the earliest in the sequence of patterns of the current reference signal listed in the new active list, after the processing of the preceding active pattern, such that the propagation of each dynamic programming path is achieved without the need to search the new active list to identify which reference patterns, of the current reference signal, have been listed on the new active list as a result of processing preceding active patterns.

29. A computer-readable medium according to claim 28, wherein said first propagating comprises:

(i) a determining instructing means for instructing the computer to determine which of the reference patterns of the current reference signal are potentially at an end of the dynamic programming path which ends at said current active pattern for said succeeding input pattern, based upon said constraints which are placed on the dynamic programming path propagation; and

(ii) a second propagating instructing means for instructing the computer to propagate the dynamic programming path, an updating instructing means for instructing the computer to update the new active list, and a copying instructing means for instructing the computer to copy appropriate cumulative values into stores associated with the reference patterns caused to be determined by the determining instructing means by processing each reference pattern caused to be determined by said determining instructing means.

30. A computer-readable medium according to claim 29, wherein each reference pattern caused to be determined by said determining instructing means, except the current active pattern if it is one of those reference patterns, is caused to be processed by the following means:

a discriminating instructing means for causing the computer to discriminate whether a current reference pattern being processed is sequentially beyond any reference pattern identified by said pointer; and

a copying instructing means for causing the computer to copy (i) the cumulative value stored in the store associated with said current active pattern into the store associated with the current reference pattern, if it is better than the cumulative value already stored there and if the current reference pattern is the same as or sequentially beyond the identified reference pattern; and

(ii) the cumulative value stored in the store associated with said current active pattern into the store associated with said current reference pattern, and to list said current reference pattern in the new active list, if the current reference pattern is sequentially before the identified reference pattern or if there are no reference patterns of the current reference signal in the new active list.

31. A computer-readable medium according to claim 29, wherein said first propagating instructing means further comprises a listing instructing means for causing the computer to list the current active pattern in the new active list, if this is not contrary to any constraint placed on the dynamic programming process if one of the reference patterns caused to be determined by said determining instructing means is the current active pattern.

32. A computer-readable medium according to claim 29, wherein said program instructing means comprises a reference pattern processing instructing means for instructing the computer to process said reference patterns in reverse sequential order.

33. A computer-readable medium according to claim 28, said program instructing means further comprising a setting instructing means for causing the computer to set said pointer associated with the current reference signal to identify the reference pattern on the new active list which is the earliest in the sequence of patterns of the current reference signal once the new active list has been updated by the processing of the current active pattern.

34. A computer-readable medium according to claim 28, said program instructing means further comprising a processing instructing means for instructing the computer to process the current input pattern before the succeeding input pattern is received.

35. A computer-readable medium according to claim 28, wherein said program instructing means further comprises a receiving code for causing the computer to receive the whole input signal before the first input pattern is processed.

36. A computer-readable medium according to claim 28, wherein said program instructing means further comprises an associating instructing means for causing the computer to associate a current active list and a new active list with each reference signal.

37. A computer-readable medium according to claim 28, wherein said first propagating instructing means cause the computer to list said reference patterns in said new active list in reverse sequential order.

38. A computer-readable medium according to claim 28, said program instructing means further comprises a preventing instructing means for instructing the computer to prevent said current active pattern from being caused to be processed by said updating and second propagating instructing means if the cumulative value associated therewith is worse than a preset threshold value.

39. A computer-readable medium according to claim 38, said program instructing means further comprising a varying instructing means for causing the computer to vary said threshold value to try to maintain the number of active patterns processed for each input pattern to be below a given maximum number.

40. A computer-readable medium according to claim 39, wherein said varying instructing means also causes the computer to set the threshold value used during the processing of the succeeding input pattern during the processing of the current input pattern, dependent upon the total number of reference patterns in the new active list or lists for the succeeding input pattern.

41. A computer-readable medium according to claim 28, said program instructing means further comprising a constraining instructing means for instructing the computer to constrain said dynamic programming path propagation so that a limit is placed on the number of consecutive input patterns that a reference pattern can be matched with, on the same dynamic programming path.

42. A computer-readable medium according to claim 28, said program instructing means further comprising a constraining instructing means for instructing the computer to constrain said dynamic programming path propagation so that if a reference pattern is matched with an input pattern, then a limited number of reference patterns which are sequentially beyond that reference pattern can be matched with the succeeding input pattern, on the same dynamic programming path.

43. A computer-readable medium according to claim 28, said program instructing means further comprising a constraining instructing means for instructing the computer to constrain said dynamic programming path propagation so that if a reference pattern is matched with an input pattern, then reference patterns which are sequentially before that reference pattern can not be matched with the succeeding input pattern, on the same dynamic programming path.

44. A computer-readable medium according to claim 28, wherein said reference patterns are representative of a template, and wherein said cumulative values are distance measures.

45. A computer-readable medium according to claim 28, wherein said reference patterns are representative of a statistical model, and wherein said cumulative values are probabilistic measures.

46. A computer-readable medium according to claim 28, wherein said program instructing means further comprises a instructing means for causing the computer to match said input signal with a sequence of said reference signals, by allowing dynamic programming paths propagating in one reference signal to subsequently propagate into other reference signals.

47. A computer-readable medium according to claim 46, wherein said program instructing means further comprises instructing means for causing the computer to cause to propagate further only the dynamic programming path which has the best cumulative value of all the dynamic programming paths which leave the reference pattern for the current input pattern.

48. A computer-readable medium according to claim 46, wherein said program instructing means further comprises a constraining instructing means for causing the computer to constrain the sequence of reference signals, against which the input signal can be matched, in accordance with a set of defined rules.

49. A computer-readable medium according to claim 46, wherein said program instructing means further comprises a processing instructing means for causing the computer to process, after the current input pattern has been processed with respect to the reference signals, those reference signals to update the new active list and to update the cumulative values associated with the reference patterns at the beginning of those reference signals, which are potentially at an end of a dynamic programming path which enters the corresponding reference signal, based upon said constraints, using the best cumulative value of all the dynamic programming paths which can enter that reference signal.

50. A computer-readable medium according to claim 28, wherein said input signal is a speech signal, and wherein each input pattern comprises a number of parameters representative of the acoustic properties of the speech signal during a corresponding time frame.

51. A computer-readable medium according to claim 50, wherein said program instructing means further comprises instructing means for causing the computer to match said input speech signal to a sequence of reference speech signals, and to define the allowed sequences by a language model.

52. A computer-readable medium according to claim 51, wherein said program instructing means further comprises instructing means for causing the computer to match the end of a dynamic programming path which leaves a reference signal when processing the current input pattern with a pattern representative of background noise for succeeding input patterns.

53. Computer executable software stored on a computer-readable medium, for instructing a computer to perform speech recognition to recognize an acoustic speech signal produced by a user and input into a microphone which produces an electrical speech signal representative thereof by comparing the electrical speech signal with a plurality of reference speech signals, the software comprising:

a first instruction for instructing the computer to extract a sequence of input patterns from the input electrical speech signal representative of the input acoustic speech signal produced by the user;

a second instruction for instructing the computer to store sequences of reference patterns, each sequence being representative of a corresponding reference speech signal;

a third instruction for instructing the computer to match the sequence of input patterns with the reference speech signal using a dynamic programming pattern matching process, the third instruction comprising a fourth instruction for instructing the computer to process a current input pattern with respect to at least some of the reference signals in turn, said fourth instruction comprising;

a fifth instruction for instructing the computer to define as active patterns the reference patterns of a current reference signal which are at the end of a dynamic programming path for the current into pattern being processed, each path representing a possible matching between an ordered sequence of reference patterns and an ordered sequence of input patterns ending at said current input pattern, and for listing the active patterns for the current input pattern in a current active list;

a sixth instruction for instructing the computer to store in a store associated with each active pattern, a cumulative value representative of a score for the dynamic programming path which ends at that active pattern for said current input pattern; and

a seventh instruction for instructing the computer to process each active pattern of said current reference signal in reverse sequential order, to update said cumulative values and to propagate said dynamic programming paths based on constraints which are placed on the dynamic programming path propagation, said seventh instruction comprising;

an eighth instruction for instructing the computer to update the cumulative value stored in the store associated with a current active pattern being processed, using said current input pattern; and

a ninth instruction for instructing the computer to propagate the dynamic programming path associated with the current active pattern, and for listing, if it is not already listed, each reference pattern of the current reference signal, which may be at the end of that dynamic programming path for the succeeding input pattern, in a new active list; and

wherein said ninth instruction instructs the computer to control the propagating means using a pointer associated with the current reference signal, which identifies the reference pattern which is the earliest in the sequence of patterns of the current reference signal listed in the new active list, after the processing of the preceding active pattern, such that the ninth instruction instructs the computer to propagate each dynamic programming path without the need to search the new active list to identify which reference patterns, of the current reference signal, have been listed on the new active list as a result of said seventh instruction instructing the computer to process preceding active patterns of the current reference signal; and

a tenth instruction for instructing the computer to display recognized speech derived from the cumulative values caused to be determined by said third instruction or to use the recognized speech as operator commands to control at least one operation of the computer from the cumulative values caused to be determined by said third instruction.

54. The software according to claim 53, wherein the ninth instruction comprises:

(i) an eleventh instruction for instructing the computer to determine which of the reference patterns of the current reference signal are potentially at an end of the dynamic programming path which ends at said current active pattern for said succeeding input pattern, based upon said constraints which are placed on the dynamic programming path propagation; and

(ii) a twelfth instruction for instructing the computer to copy appropriate cumulative values into stores associated with the reference patterns caused to be determined by said eleventh instruction.

55. The software according to claim 54, wherein said ninth instruction further comprises a thirteen instruction for instructing the computer to process each of the reference patterns caused to be determined by said eleventh instruction comprising:

(a) a fourteenth instruction for instructing the computer to discriminate whether a current reference pattern being processed is sequentially beyond any reference pattern identified by said pointer; and

(b) a fifteenth instruction, responsive to said discriminating caused to be performed by said fourteenth instruction code!, for instructing the computer to (i) copy the cumulative value stored in the store associated with said current active pattern into the store associated with the current reference pattern, if it is better than the cumulative value already stored there and if the current reference pattern is the same as or sequentially beyond the identified reference pattern; and

to (ii) copy the cumulative value stored in the store associated with said current active pattern into the store associated with said current reference pattern, and listing said current reference pattern in the new active list, if the current reference pattern is sequentially before the identified reference pattern or if there are no reference patterns of the current reference signal in the new active list.

56. The software according to claim 55, wherein said thirteenth instruction instructs the computer to process said reference patterns caused to be determined by said eleventh instruction in reverse sequential order.

57. The software according to claim 54, wherein said ninth instruction also instructs the computer to list the current active pattern in the new active list, if this is not contrary to any constraint placed on the dynamic programming process if one of the reference patterns caused to be determined by said eleventh instruction is the current active pattern.

58. The software according to claim 53, further comprising means an eleventh instruction for instructing the computer to set the pointer associated with the current reference signal to identify the reference pattern on the new active list which is the earliest in the sequence of patterns of the current reference signal once the new active list has been updated by the processing of the current active pattern instructed to be performed by said fourth instruction.

59. The software according to claim 53, wherein the fourth instruction instructs the computer to process the current input pattern before the succeeding input pattern is received.

60. The software according to claim 53, wherein the whole electrical speech signal is input before said fourth instruction instructs the computer to process the first input pattern.

61. The software according to claim 53, wherein said fifth and ninth instructions instruct the computer to generate a current active list and a new active list for each reference signal.

62. The software according to claim 53, wherein said ninth instruction instructs the computer to list said reference patterns in said new active list in reverse sequential order.

63. The software according to claim 53, further comprising an eleventh instruction for instructing the computer to prevent said seventh instruction from instructing the computer to process the current active pattern, if the cumulative value is associated therewith is worse than a preset threshold value.

64. The software according to claim 63, further comprising a twelfth instruction for instructing the computer to vary said threshold value in order to try to maintain the number of active patterns caused to be processed by said seventh instruction for each input pattern caused to be processed by said fourth instruction to be below a given maximum number.

65. The software according to claim 64, wherein the twelfth instruction code instructs the computer to set the threshold value used during the processing of the succeeding input pattern during the processing of the current input pattern, and is dependent upon the total number of reference patterns in the new active list or lists for the succeeding input pattern.

66. The software according to claim 53, further comprising an eleventh instruction for instructing the computer to constrain said dynamic programming path propagation so that a limit is placed on the number of consecutive input patterns that a reference pattern can be matched with, on the same dynamic programming path.

67. The software according to claim 53, further comprising an eleventh instruction for instructing the computer to constrain said dynamic programming path propagation so that if a reference pattern is matched with an input pattern, then a limited number of reference patterns which are sequentially beyond that reference pattern can be matched with the succeeding input pattern, on the same dynamic programming path.

68. The software according to claim 53, further comprising an eleventh instruction for instructing the computer to constrain said dynamic programming path propagation so that if a reference pattern is matched with an input pattern, then reference patterns which are sequentially before that reference pattern can not be matched with the succeeding input pattern, on the same dynamic programming path.

69. The software according to claim 53, wherein said reference patterns are representative of a template, and wherein said cumulative values are distance measures.

70. The software according to claim 53, wherein said reference patterns are representative of a statistical model, and wherein said cumulative values are probabilistic measures.

71. The software according to claim 53, further comprising an eleventh instruction for instructing the computer to match said input electrical speech signal with a sequence of said reference signals, by allowing dynamic programming paths propagating in one reference signal to subsequently propagate into other reference signals.

72. The software according to claim 71, further comprising a twelfth instruction for instructing the computer to propagate further only the dynamic programming path, which has the best cumulative value of all the dynamic programming paths which leave the reference pattern for the current input pattern.

73. The software according to claim 71, further comprising a twelfth instruction for instructing the computer to use a set of defined rules for constraining the sequence of reference signals against which the input signal can be matched.

74. The software according to claim 71, further comprising a twelfth instruction for instructing the computer to, after the current input pattern has been processed with respect to the reference signals, process those reference signals to update the new active list and to update the cumulative values associated with the reference patterns at the beginning of those reference signals, which are potentially at an end of a dynamic programming path which enters the corresponding reference signal, based upon said constraints, using the best cumulative value of all the dynamic programming paths which can enter that reference signal.

75. The software according to claim 53, wherein each input pattern comprises a number of parameters representative of the acoustic properties of the speech signal during a corresponding time frame.

76. The software according to claim 75, wherein the allowed sequences are defined by a language model.

77. The software according to claim 76, further comprising a twelfth instruction for instructing the computer match the end of a dynamic programming path which leaves a reference signal when processing the current input pattern with a pattern representative of background noise for succeeding input patterns.

78. A computer-readable memory medium for storing a first computer-readable program embodied therein for instructing a computer to perform a speech recognition process for recognizing an input speech signal by comparing it with a plurality of reference speech signals, the first program comprising:

an extracting instructing means for causing the computer to extract a sequence of input patterns representative of said input speech signal;

a first storing instructing means for causing the computer to store sequences of reference patterns, each sequence being representative of a corresponding reference speech signal;

a matching instructing means for causing the computer to match the sequence of input patterns with the reference speech signals, the matching instructing means comprising;

a second program instructing means embodied for instructing a computer to perform a dynamic programming pattern matching process between a sequence of input patterns representative of an input signal, and a number of stored sequences of reference patterns, each sequence being representative of a reference signal, wherein the second program instructing means instructs the computer to process each input pattern in turn with respect to at least some of the reference signals, said second program instructing means comprising;

(1) a defining instructing means for instructing the computer to define as active patterns the reference patterns of a current reference signal which are at the end of a dynamic programming path for a current input pattern being processed, each path representing a possible matching between an ordered sequence of reference patterns and an ordered sequence of input patterns ending at said current input pattern, and for instructing the computer to list the active patterns for the current input pattern in a current active list;

(2) a second storing instructing means for instructing the computer to, for each active pattern, store in a store associated with that active pattern, a cumulative value representative of a score for the dynamic programming path which ends at that active pattern for said current input pattern; and

(3) an updating and propagating instructing means for instructing the computer to update said cumulative values and to propagate said dynamic programming paths based on constraints which are placed on the dynamic programming path propagation, by processing each active pattern of said current reference signal in reverse sequential order, comprising;

(A) an updating instructing means for instructing the computer to update the cumulative value stored in the store associated with a current active pattern being processed, using said current input pattern; and

then

(B) a first propagating instructing means for instructing the computer to propagate the dynamic programming path associated with the current active pattern, and to list, if it is not already listed, each reference pattern of the current reference signal, which may be at the end of that dynamic programming path for the succeeding input pattern, in a new active list; and

wherein the propagating operation for each dynamic programming path for a current reference signal is controlled using a pointer associated with the current reference signal, which identifies the reference pattern which is the earliest in the sequence of patterns of the current reference signal listed in the new active list, after the processing of the preceding active pattern, such that the propagation of each dynamic programming path is achieved without the need to search the new active list to identify which reference patterns, of the current reference signal, have been listed on the new active list as a result of processing preceding active patterns; and

a providing instructing means for causing the computer to provide a recognition result from cumulative values determined by said matching step.

79. A computer-readable memory according to claim 78, wherein said providing instructing means comprises instructing means for causing the computer to provide said recognition result by determining which of the dynamic programming paths which end at the last input pattern in the sequence has the best cumulative value.

80. A speech recognition apparatus for recognizing all input speech signal by comparing it with a plurality of reference speech signals, the apparatus comprising:

a first circuit configured to extract a sequence of input patterns representative of an input speech signal from an input speech signal;

a second circuit configured to store sequences of reference patterns, each sequence being representative of a corresponding reference speech signal;

a third circuit configured to match the sequence of input patterns with the reference speech signal, using a fourth circuit configured to perform a dynamic programming pattern matching process, the fourth circuit comprising;

a fifth circuit configured to input the sequence of input patterns representative of an input speech signal; and

a sixth circuit configured to process a current input pattern with respect to at least some of the reference signals in turn, comprising;

(1) a seventh circuit configured to define as active patterns the reference patterns of a current reference signal which are at the end of a dynamic programming path for the current input pattern being processed, each path representing a possible matching between an ordered sequence of reference patterns and an ordered sequence of input patterns ending at said current input pattern, and for listing the active patterns for the current input pattern in a current active list;

(2) an eighth circuit configured to store in a storeassociated with each active pattern, a cumulative value representative of a score for the dynamic programming path which ends at that active pattern for said current input pattern; and

(3) a ninth circuit configured to process each active pattern of said current reference signal in reverse sequential order, to update said cumulative values and to propagate said dynamic programming paths based on constraints which are placed on the dynamic programming path propagation, ninth circuit comprising;

(a) a tenth circuit configured to update the cumulative value stored in the store associated with a current active pattern being processed, using said current input pattern; and

(b) an eleventh circuit configured to propagate the dynamic programming path associated with the current active pattern, and for listing, if it is not already listed, each reference pattern of the current reference signal, which may be at the end of that dynamic programming path for the succeeding input pattern, in a new active list; and

wherein said eleventh circuit is controlled using a pointer associated with the current reference signal, which identifies the reference pattern which is the earliest in the sequence of patterns of the current reference signal listed in the new active list, after the processing of the preceding active pattern, such that the propagating means can propagate each dynamic programming path without the need to search the new active list to identify which reference patterns, of the current reference signal, have been listed on the new active list as a result of said ninth circuit processing preceding active patterns of the current reference signal, and wherein said input signal is a speech signal, and wherein each input pattern comprises a number of parameters representative of the acoustic properties of the speech signal during a corresponding time frame; and

a twelfth circuit configured to provide a recognition result from cumulative values determined by said third circuit.

81. An apparatus according to claim 80, wherein said twelfth circuit is operable to provide said recognition result by determining which of the dynamic programming paths which end at the last input pattern in the sequence has the best cumulative value.

82. An apparatus for performing a dynamic programming pattern matching process, the apparatus comprising:

a first circuit configured to input a sequence of input patterns representative of an input signal;

a second circuit configured to store a plurality of sequences of references patterns, each sequence being representative of a reference signal;

a third circuit configured to process a current input pattern with respect to at least some of the reference signals in turn, comprising;

(1) a fourth circuit configured to define as active patterns the reference patterns of a current reference signal which are at the end of a dynamic programming path for the current into pattern being processed, each path representing a possible matching between an ordered sequence of reference patterns and all ordered sequence of input patterns ending at said current input pattern, and for listing the active patterns for the current input pattern in a current active list;

(2) a fifth circuit configured to store in a store associated with each active pattern, a cumulative value representative of a score for thc dynamic programming path which ends at that active pattern for said current input pattern; and

(3) a sixth circuit configured to process each active pattern of said current reference signal in reverse sequential order, to update said cumulative values and to propagate said dynamic programming paths based on constraints which are placed on the dynamic programming path propagation, said sixth circuit comprising;

(a) a seventh circuit configured to update the cumulative value stored in the store associated with a current active pattern being processed, using said current input pattern; and

(b) an eighth circuit configured to propagate the dynamic programming path associated with the current active pattern, and for listing, if it is not already listed, each reference pattern of the current reference signal, which may be at the end of that dynamic programming path fur the succeeding input pattern, in a new active list; and

wherein said eighth circuit is controlled using a pointer associated with the current reference signal, which identifies the reference pattern which is the earliest iii the sequence of patterns of the current reference signal listed in the new active list, after the processing of the preceding active pattern, such that the eighth circuit can propagate each dynamic programming path without the need to search the new active list to identify which reference patterns, of the current reference signal, have been listed on the new active list as a result of said sixth circuit processing preceding active patterns of the current reference signal.

83. An apparatus according to claim 82, wherein the eighth circuit comprises:

(i) a ninth circuit configured to determine which of the reference patterns of the current reference signal are potentially at an end of the dynamic programming path which ends at said current active pattern for said succeeding input pattern, based upon said constraints which are placed on the dynamic programming path propagation; and

(ii) a tenth circuit configured to copy appropriate cumulative values into stores associated with the reference patterns determined by said ninth circuit.

84. An apparatus according to claim 83, wherein said ninth circuit further comprises an eleventh circuit configured to process each of the reference patterns determined by said ninth circuit comprising:

(a) a twelfth circuit configured to discriminate whether a current reference pattern being processed is sequentially beyond any reference pattern identified by said pointer; and

(b) a thirteenth circuit, responsive to said eleventh circuit, configured to (i) copy the cumulative value stored in the store associated with said current active pattern into the store associated with the current reference pattern, if it is better than the cumulative value already stored there and if the current reference pattern is the same as or sequentially beyond the identified reference pattern; and

(ii) copy the cumulative value stored in the store associated with said current active pattern into the store associated with said current reference pattern, and listing said current reference pattern in the new active list, if the current reference pattern is sequentially before the identified reference pattern or if there are no reference patterns of the current reference signal in the new active list.

85. An apparatus according to claim 84, wherein said twelfth circuit is configured to process said reference patterns determined by said ninth circuit in reverse sequential order.

86. An apparatus according to claim 83, wherein said eighth circuit is operable to list the current active pattern in the new active list, if this is not contrary to any constraint placed on the dynamic programming process if one of the reference patterns determined by said ninth circuit is the current active pattern.

87. An apparatus according to claim 82, further comprising a ninth circuit configured to set the pointer associated with thc current reference signal to identify the reference pattern on the new active list which is the earliest in the sequence of patterns of the current reference signal once the new active list has been updated by the processing of the current active pattern by said sixth circuit.

88. An apparatus according to claim 82, wherein the first circuit is configured to process the current input pattern before the succeeding input pattern is received.

89. An apparatus according to claim 82, wherein said first circuit is configured to input the whole input signal before the first input pattern is processed by said third circuit.

90. An apparatus according to claim 82, further comprising a current active list and a new active list for each reference signal.

91. An apparatus according to claim 82, wherein said eighth circuit is configured to list said reference patterns in said new active list in reverse sequential order.

92. An apparatus according to claim 82, further comprising a ninth circuit configured to prevent said sixth circuit from processing the current active pattern, if the cumulative value is associated therewith is worse than a preset threshold value.

93. An apparatus according to claim 92, further comprising an tenth circuit configured to vary said threshold value in order to try to maintain the number of active patterns processed by said sixth circuit for each input pattern processed by said third circuit to be below a given maximum number.

94. An apparatus according to claim 93, wherein the tenth circuit is configured to set the threshold value used during the processing of the succeeding input pattern during the processing of the current input pattern, and is dependent upon the total number of reference patterns in the new active list or lists for the succeeding input pattern.

95. An apparatus according to claim 82, further comprising a ninth circuit configured to constrain said dynamic programming path propagation so that a limit is placed on the number of consecutive input patterns that a reference pattern can be matched with, on the same dynamic programming path.

96. An apparatus according to claim 82, further comprising a ninth circuit configured to constrain said dynamic programming path propagation so that if a reference pattern is matched with an input pattern, then a limited number of reference patterns which are sequentially beyond that reference pattern can be matched with the succeeding input pattern, on the same dynamic programming path.

97. An apparatus according to claim 82, further comprising a ninth circuit configured to constrain said dynamic programming path propagation so that if a reference pattern is matched with an input pattern, then reference patterns which are sequentially before that reference pattern can not be matched with the succeeding input pattern, on the same dynamic programming path.

98. An apparatus according to claim 82, wherein said reference patterns are representative of a template, and wherein said cumulative values are distance measures.

99. An apparatus according to claim 82, wherein said reference patterns are representative of a statistical model, and wherein said cumulative values are probabilistic measures.

100. An apparatus according to claim 82, wherein said apparatus is arranged so that said input signal can be matched with a sequence of said reference signals, by allowing dynamic programming paths propagating in one reference signal to subsequently propagate into other reference signals.

101. An apparatus according to claim 100, wherein said apparatus is arranged so that only the dynamic programming path, which has the best cumulative value of all the dynamic programming paths which leave the reference pattern for the current input pattern, propagates further.

102. An apparatus according to claim 100, further comprising a ninth circuit configured to use a set of defined rules for constraining the sequence of reference signals against which the input signal can be matched.

103. An apparatus according to claim 100, further comprising a ninth circuit configured to, after the current input pattern has been processed with respect to the reference signals by said third circuit, to process those reference signals to update the new active list and to update the cumulative values associated with the reference patterns at the beginning of those reference signals, which are potentially at an end of a dynamic programming path which enters the corresponding reference signal, based upon said constraints, using the best cumulative value of all the dynamic programming paths which can enter that reference signal.

104. An apparatus according to claim 82, wherein said input signal is a speech signal, and wherein each input pattern comprises a number of parameters representative of the acoustic properties of the speech signal during a corresponding time frame.

105. An apparatus according to claim 104, wherein said input speech signal can be matched to a sequence of reference speech signals, and wherein the allowed sequences are defined by a language model.

106. An apparatus according to claim 105, wherein said apparatus is arranged so that the end of a dynamic programming path which leaves a reference signal when processing the current input pattern can be matched with a pattern representative of background noise for succeeding input patterns.

107. A method of performing a dynamic programming pattern matching process between a sequence of input patterns representative of an input signal, and a number of stored sequences of reference patterns, each sequence being representative of a reference signal, wherein the method processes each input pattern in turn with respect to at least some of the reference signals in turn, by:

(1) defining as active patterns the reference patterns of a current reference signal which are at the end of a dynamic programming path for a current input pattern being processed, each path representing a possible matching between an ordered sequence of reference patterns and an ordered sequence of input patterns ending at said current input pattern, and listing the active patterns for the current input pattern in a current active list;

(2) for each active pattern, storing in a store associated with that active pattern, a cumulative value representative of the score for the dynamic programming path which ends at that active pattern for said current input pattern; and

(3) updating said cumulative values and propagating said dynamic programming paths based on constraints which are placed on the dynamic programming path propagation, by processing each active pattern of said current reference signal in reverse sequential order, by;

(A) updating the cumulative value stored in the store associated with a current active pattern being processed, using said current input pattern; and

then

(B) propagating the dynamic programming path associated with the current active pattern, and listing, if it is not already listed, each reference pattern of the current reference signal, which may be at the end of that dynamic programming path for the succeeding input pattern, in a new active list; and

wherein the propagation of each dynamic programming path for a current reference signal is controlled using a pointer associated with the current reference signal, which identifies the reference pattern which is the earliest in the sequence of patterns of the current reference signal listed in the new active list, after the processing of the preceding active pattern, such that the propagation of each dynamic programming path is achieved without the need to search the new active list to identify which reference patterns, of the current reference signal, have been listed on the new active list as a result of processing preceding active patterns.

108. A method according to claim 107, wherein the following steps are performed in step (B):

(i) determining which of the reference patterns of the current reference signal are potentially at an end of the dynamic programming path which ends at said current active pattern for said succeeding input pattern, based upon said constraints which are placed on the dynamic programming path propagation; and

then

(ii) propagating the dynamic programming path, updating the new active list and copying appropriate cumulative values into stores associated with the reference patterns determined in step (i), by processing each reference pattern determined in step (i).

109. A method according to claim 108, wherein each reference pattern determined in step (i) of claim 2, except the current active pattern if it is one of those reference patterns, is processed by:

(a) determining whether a current reference pattern being processed is sequentially beyond any reference pattern identified by said pointer;

then

(b) if the current reference pattern is the same as or sequentially beyond the identified reference pattern, then copying the cumulative value stored in the store associated with said current active pattern into the store associated with the current reference pattern, if it is better than the cumulative value already stored there; and

(c) if the current reference pattern is sequentially before the identified reference pattern or if there are no reference patterns of the current reference signal in the new active list, then copying the cumulative value stored in the store associated with said current active pattern into the store associated with said current reference pattern, and listing said current reference pattern in the new active list.

110. A method according to claim 108, wherein if one of the reference patterns determined in step (i) of claim 2 is the current active pattern, then the processing of the current active pattern in step (ii) of claim 2 comprises the step of listing the current active pattern in the new active list, if this is not contrary to any constraint placed on the dynamic programming process.

111. A method according to claim 108, wherein the reference patterns processed in step (ii) of claim 2, are processed in reverse sequential order.

112. A method according to claim 107, wherein after the new active list has been updated by the processing of the current active pattern, said pointer associated with the current reference signal is set to identify the reference pattern on the new active list which is the earliest in the sequence of patterns of the current reference signal.

113. A method according to claim 107, wherein the processing of the current input pattern is performed before the succeeding input pattern is received.

114. A method according to claim 107, wherein the whole input signal is received before the first input pattern is processed.

115. A method according to claim 107, wherein each reference signal has associated therewith a current active list and new active list.

116. A method according to claim 107, wherein when a reference pattern is added to said new active list it is placed at the end thereof.

120. A method according to claim 107, wherein one of said dynamic programming constraints is that a limit is placed on the number of consecutive input patterns that a reference pattern can be matched with, on the same dynamic programming path.

121. A method according to claim 107, wherein one of said dynamic programming constraints is that if a reference pattern is matched with an input pattern, then a limited number of reference patterns which are sequentially beyond that reference pattern can be matched with the succeeding input pattern, on the same dynamic programming path.

122. A method according to claim 107, wherein one of said dynamic programming constraints is that if a reference pattern is matched with an input pattern, then reference patterns which are sequentially before that reference pattern can not be matched with the succeeding input pattern, on the same dynamic programming path.

123. A method according to claim 107, wherein said reference patterns are representative of a template, and wherein said cumulative values are distance measures.

124. A method according to claim 107, wherein said reference patterns are representative of a statistical model, and wherein said cumulative values are probabilistic measures.

125. A method according to claim 107, wherein said input signal can be matched with a sequence of said reference signals, by allowing dynamic programming paths propagating in one reference signal to subsequently propagate into other reference signals.

126. A method according to claim 125, wherein only the dynamic programming path which has the best cumulative value of all the dynamic programming paths which leaves the reference pattern for the current input pattern, propagates further.

127. A method according to claim 125, wherein the sequence of reference signals, against which the input signal can be matched, is constrained by a set of defined rules.

128. A method according to claim 125, wherein after the current input pattern has been processed with respect to the reference signals, the method processes those reference signals to update the new active list and to update the cumulative values associated with the reference patterns at the beginning of those reference signals, which are potentially at an end of a dynamic programming path which enters the corresponding reference signal, based upon said constraints, using the best cumulative value of all the dynamic programming paths which can enter that reference signal.

129. A method according to claim 107, wherein said input signal is representative of a speech signal, and wherein each input pattern comprises a number of parameters representative of the acoustic properties of the speech signal during a corresponding time frame.

130. A method according to claim 129, wherein said input speech signal can be matched to a sequence of reference speech signals, and wherein the allowed sequences are defined by a language model.

131. A method according to claim 130, wherein the end of a dynamic programming path which leaves a reference signal when processing the current input pattern can be matched with a pattern representative of background noise for succeeding input patterns.

132. A speech recognition method for recognising an input speech signal by comparing it with a plurality of reference speech signals, the method comprising the steps of:

extracting a sequence of input patterns representative of said input speech signal;

storing sequences of reference patterns, each sequence being representative of a corresponding reference speech signal;

matching the sequence of input patterns with the reference speech signals, using a method according to claim 129; and

providing a recognition result from cumulative values determined by said matching step.

133. A method according to claim 132, wherein said recognition result is provided by determining which of the dynamic programming paths which end at the last input pattern in the sequence has the best cumulative value.