Unsupervised data-driven pronunciation modeling

• US 7,165,032 B2
• Filed: 11/22/2002
• Issued: 01/16/2007
• Est. Priority Date: 09/13/2002
• Status: Expired due to Fees

• Current Assignee: Apple Inc.
• Original Assignee: Apple Computer Incorporated (Apple Inc.)
• Inventors: Bellegarda, Jerome R.
• Primary Examiner(s): ABEBE, DANIEL DEMELASH
• Application Number:
  US10/302,438
• Publication Number:
  US 20040054533A1
• Time in Patent Office: 1,516 Days
• Field of Search: 704/258, 704/260
• US Class Current: 704/258
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams

Abstract

Pronunciation for an input word is modeled by generating a set of candidate phoneme strings having pronunciations close to the input word in an orthographic space. Phoneme sub-strings in the set are selected as the pronunciation. In one aspect, a first closeness measure between phoneme strings for words chosen from a dictionary and contexts within the input word is used to determine the candidate phoneme strings. The words are chosen from the dictionary based on a second closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary. In another aspect, the phoneme sub-strings are selected by aligning the candidate phoneme strings on common phoneme sub-strings to produce an occurrence count, which is used to choose the phoneme sub-strings for the pronunciation.

Claims

1. A computerized method comprising:

receiving, by a client from a server, an orthographic space;

generating a set of candidate phoneme strings having pronunciations close to an input word in the orthographic space; and

selecting phoneme sub-strings from the set as a pronunciation for the input word.

2. The computerized method of claim 1, wherein generating a set of candidate phoneme strings comprises:

determining phoneme strings for the set based on a closeness measure between phoneme strings for words chosen from a dictionary and contexts within the input word.

3. The computerized method of claim 2 further comprising:

choosing the words from the dictionary based on a closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary.

4. The computerized method of claim 3 further comprising:

mapping words in a dictionary and a set of letter patterns that are characteristic of the words into the orthographic space to create the orthographic anchors.

5. The computerized method of claim 4, wherein mapping words comprises:

creating a matrix of feature vectors from the letter patterns and the words;

decomposing the matrix into row vectors and column vectors; and

selecting vectors from one of the row vectors and column vectors as the orthographic anchors.

6. The computerized method of claim 1, wherein selecting phoneme sub-strings comprises:

aligning the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts; and

choosing the phoneme sub-strings for the pronunciation based on the occurrence counts.

7. A computerized method comprising:

receiving, by a server from a client, an input word and a neighborhood that identifies phoneme strings having pronunciations close to the input word in an orthographic space;

generating a set of candidate phoneme strings for the input word from the phoneme strings;

selecting phoneme sub-strings from the set as a pronunciation for the input word; and

returning the pronunciation to the client.

8. The computerized method of claim 7, wherein generating a set of candidate phoneme strings comprises:

determining phoneme strings for the set based on a closeness measure between the phoneme strings and contexts within the input word.

9. The computerized method of claim 7, wherein selecting phoneme sub-strings comprises:

aligning the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts; and

choosing the phoneme sub-strings for the pronunciation based on the occurrence counts.

10. The computerized method of claim 7, wherein the neighborhood is an orthographic neighborhood and further comprising:

creating, by the server, a pronunciation neighborhood from phoneme strings for words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

11. The computerized method of claim 10 further comprising:

choosing, by the client, words for the orthographic neighborhood from a dictionary based on a closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary.

12. The computerized method of claim 11 further comprising:

mapping the words in the dictionary and a set of letter patterns that are characteristic of the words into the orthographic space to create the orthographic anchors.

13. The computerized method of claim 12, wherein mapping the words comprises:

creating a matrix of feature vectors from the letter patterns and the words;

decomposing the matrix into row vectors and column vectors; and

selecting vectors from one of the row vectors and column vectors as the orthographic anchors.

14. The computerized method of claim 7, wherein the neighborhood is a pronunciation neighborhood and further comprising:

creating, by the client, the pronunciation neighborhood from the phoneme strings for words in an orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

15. A computerized method comprising:

receiving, by a client from a server, a neighborhood that identifies phoneme strings having pronunciations close to an input word in an orthographic space;

generating a set of candidate phoneme strings for the input word from the phoneme strings; and

selecting phoneme sub-strings from the set as a pronunciation for the input word.

16. The computerized method of claim 15, wherein generating a set of candidate phoneme strings comprises:

determining phoneme strings for the set based on a closeness measure between the phoneme strings and contexts within the input word.

17. The computerized method of claim 15, wherein selecting phoneme sub-strings comprises:

aligning the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts; and

choosing the phoneme sub-strings for the pronunciation based on the occurrence counts.

18. The computerized method of claim 15, wherein the neighborhood is an orthographic neighborhood and further comprising:

creating, by the client, a pronunciation neighborhood from phoneme strings for words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

19. The computerized method of claim 18 further comprising:

choosing, by the server, words for the orthographic neighborhood from a dictionary based on a closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary.

20. The computerized method of claim 19 further comprising:

mapping the words in the dictionary and a set of letter patterns that are characteristic of the words into the orthographic space to create the orthographic anchors.

21. The computerized method of claim 20, wherein mapping the words comprises:

creating a matrix of feature vectors from the letter patterns and the words;

decomposing the matrix into row vectors and column vectors; and

selecting vectors from one of the row vectors and column vectors as the orthographic anchors.

22. The computerized method of claim 15, wherein the neighborhood is a pronunciation neighborhood and further comprising:

creating, by the server, the pronunciation neighborhood from the phoneme strings for words in an orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

23. A machine-readable medium having executable instructions to cause a machine to perform a method comprising:

receiving, by a client from a server, an orthographic space;

generating a set of candidate phoneme strings having pronunciations close to an input word in the orthographic space; and

selecting phoneme sub-strings from the set as a pronunciation for the input word.

24. The machine-readable medium of claim 23, wherein generating a set of candidate phoneme strings comprises:

determining phoneme strings for the set based on a closeness measure between phoneme strings for words chosen from a dictionary and contexts within the input word.

25. The machine-readable medium of claim 24, wherein the method further comprises:

choosing the words from the dictionary based on a closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary.

26. The machine-readable medium of claim 25, wherein the method further comprises:

mapping words in a dictionary and a set of letter patterns that are characteristic of the words into the orthographic space to create the orthographic anchors.

27. The machine-readable medium of claim 26, wherein mapping words comprises:

creating a matrix of feature vectors from the letter patterns and the words;

decomposing the matrix into row vectors and column vectors; and

selecting vectors from one of the row vectors and column vectors as the orthographic anchors.

28. The machine-readable medium of claim 23, wherein selecting phoneme sub-strings comprises:

aligning the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts; and

choosing the phoneme sub-strings for the pronunciation based on the occurrence counts.

29. A machine-readable medium having executable instructions to cause a machine to perform a method comprising:

receiving, by a server from a client, an input word and a neighborhood that identifies phoneme strings having pronunciations close to the input word in an orthographic space;

generating a set of candidate phoneme strings for the input word from the phoneme strings;

selecting phoneme sub-strings from the set as a pronunciation for the input word; and

returning the pronunciation to the client.

30. The machine-readable medium of claim 29, wherein generating a set of candidate phoneme strings comprises:

determining phoneme strings for the set based on a closeness measure between the phoneme strings and contexts within the input word.

31. The machine-readable medium of claim 29, wherein selecting phoneme sub-strings comprises:

aligning the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts; and

choosing the phoneme sub-strings for the pronunciation based on the occurrence counts.

32. The machine-readable medium of claim 29, wherein the neighborhood is an orthographic neighborhood and the method further comprises:

creating, by the server, a pronunciation neighborhood from phoneme strings for words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

33. The machine-readable medium of claim 32, wherein the method further comprises:

choosing, by the client, words for the orthographic neighborhood from a dictionary based on a closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary.

34. The machine-readable medium of claim 33, wherein the method further comprises:

mapping the words in the dictionary and a set of letter patterns that are characteristic of the words into the orthographic space to create the orthographic anchors.

35. The machine-readable medium of claim 34, wherein mapping the words comprises:

creating a matrix of feature vectors from the letter patterns and the words;

decomposing the matrix into row vectors and column vectors; and

selecting vectors from one of the row vectors and column vectors as the orthographic anchors.

36. The machine-readable medium of claim 29, wherein the neighborhood is a pronunciation neighborhood and the method further comprises:

creating, by the client, the pronunciation neighborhood from the phoneme strings for words in an orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

37. A machine-readable medium having executable instructions to cause a machine to perform a method comprising:

receiving, by a client from a server, a neighborhood that identifies phoneme strings having pronunciations close to an input word in an orthographic space;

generating a set of candidate phoneme strings for the input word from the phoneme strings; and

selecting phoneme sub-strings from the set as a pronunciation for the input word.

38. The machine-readable medium of claim 37, wherein generating a set of candidate phoneme strings comprises:

determining phoneme strings for the set based on a closeness measure between the phoneme strings and contexts within the input word.

39. The machine-readable medium of claim 37, wherein selecting phoneme sub-strings comprises:

aligning the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts; and

choosing the phoneme sub-strings for the pronunciation based on the occurrence counts.

40. The machine-readable medium of claim 37, wherein the neighborhood is an orthographic neighborhood and the method further comprises:

creating, by the client, a pronunciation neighborhood from phoneme strings for words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

41. The machine-readable medium of claim 40, wherein the method further comprises:

choosing, by the server, words for the orthographic neighborhood from a dictionary based on a closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary.

42. The machine-readable medium of claim 41, wherein the method further comprises:

mapping the words in the dictionary and a set of letter patterns that are characteristic of the words into the orthographic space to create the orthographic anchors.

43. The machine-readable medium of claim 42, wherein mapping the words comprises:

creating a matrix of feature vectors from the letter patterns and the words;

decomposing the matrix into row vectors and column vectors; and

selecting vectors from one of the row vectors and column vectors as the orthographic anchors.

44. The machine-readable medium of claim 37, wherein the neighborhood is a pronunciation neighborhood and the method further comprises:

creating, by the server, the pronunciation neighborhood from the phoneme strings for words in an orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

45. An apparatus comprising:

means for receiving, by a client from a server, an orthographic space;

means for generating a set of candidate phoneme strings having pronunciations close to an input word in the orthographic space; and

means for selecting phoneme sub-strings from the set as a pronunciation for the input word.

46. The apparatus of claim 45, wherein the means for generating comprises:

means for measuring a closeness between phoneme strings for words chosen from a dictionary and contexts within the input word to determine phoneme strings for the set.

47. The apparatus of claim 46 further comprising:

means for measuring a closeness between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary to determine word choices.

48. The apparatus of claim 47 further comprising:

means for mapping words in a dictionary and a set of letter patterns that are characteristic of the words into the orthographic space to create the orthographic anchors.

49. The apparatus of claim 48, wherein the means for mapping comprises:

means for creating a matrix of feature vectors from the letter patterns and the words;

means for decomposing the matrix into row vectors and column vectors; and

means for selecting vectors from one of the row vectors and column vectors as the orthographic anchors.

50. The apparatus of claim 45, wherein the means for selecting comprises:

means for aligning the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts; and

means for choosing the phoneme sub-strings for the pronunciation based on the occurrence counts.

51. An apparatus comprising:

means for receiving, by a server from a client, an input word and a neighborhood that identifies phoneme strings having pronunciations dose to the input word in an orthographic space;

means for generating a set of candidate phoneme strings for the input word from the phoneme strings;

means for selecting phoneme sub-strings from the set as a pronunciation for the input word; and

means for returning the pronunciation to the client.

52. The apparatus of claim 51, wherein the means for generating comprises:

means for measuring a closeness between phoneme strings for words chosen from a dictionary and contects within the input word to determine phoneme strings for the set.

53. The apparatus of claim 51, wherein the means for selecting comprises:

means for aligning the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts; and

means for choosing the phoneme sub-strings for the pronunciation based on the occurrence counts.

54. The apparatus of claim 51, wherein the neighborhood is an orthographic neighborhood and further comprising:

means for creating, by the server, a pronunciation neighborhood from phoneme strings for words in the orthographic neighborhoods the candidate phoneme strings being generated from the pronunciation neighborhood.

55. The apparatus of claim 54 further comprising:

means for measuring, by the client, a closeness between a representation of the input word in the orthographic space and orthographic anchors corresponding to words in a dictionary to choose words for the orthographic neighborhood.

56. The apparatus of claim 55 further comprising:

means for mapping the words in the dictionary and a set of letter patterns that are characteristic of the words into the orthographic space to create the orthographic anchors.

57. The apparatus of claim 56, wherein the means for mapping comprises:

means for creating a matrix of feature vectors from the letter patterns and the words;

means for decomposing the matrix into row vectors and column vectors; and

means for selecting vectors from one of the row vectors and column vectors as the orthographic anchors.

58. The apparatus of claim 51, wherein the neighborhood is a pronunciation neighborhood and further comprising:

means for creating, by the client, the pronunciation neighborhood from the phoneme strings for words in an orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

59. An apparatus comprising:

means for receiving, by a client from a server, a neighborhood that identifies phoneme strings having pronunciations dose to an input word in an orthographic space;

means for generating a set of candidate phoneme strings for the input word from the phoneme strings; and

means for selecting phoneme sub-strings from the set as a pronunciation for the input word.

60. The apparatus of claim 59, wherein the means for generating comprises:

means for measuring a closeness between phoneme strings for words chosen from a dictionary and contexts within the input word to determine phoneme strings for the set.

61. The apparatus of claim 59, wherein the means for selecting comprises:

means for aligning the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts; and

means for choosing the phoneme sub-strings for the pronunciation based on the occurrence counts.

62. The apparatus of claim 59, wherein the neighborhood is an orthographic neighborhood and further comprising:

means for creating, by the client, a pronunciation neighborhood from phoneme strings for words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

63. The apparatus of claim 62 further comprising:

means for measuring, by the server, a closeness between a representation of the input word in the orthographic space and orthographic anchors corresponding to words in a dictionary to choose words for the orthographic neighborhood.

64. The apparatus of claim 63 further comprising:

means for mapping the words in the dictionary and a set of letter patterns that are characteristic of the words into the orthographic space to create the orthographic anchors.

65. The apparatus of claim 64, wherein the means for mapping comprises:

means for creating a matrix of feature vectors from the letter patterns and the words;

means for decomposing the matrix into row vectors and column vectors; and

means for selecting vectors from one of the row vectors and column vectors as the orthographic anchors.

66. The apparatus of claim 59, wherein the neighborhood is a pronunciation neighborhood and further comprising:

means for creating, by the server, the pronunciation neighborhood from the phoneme strings for words in an orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

67. A system comprising:

a processor coupled to a memory through a bus;

an interface coupled to the processor through the bus to receive an input word;

a network interface coupled to the processor through the bus to communicate with a server; and

a client process executed from the memory by the processor to cause the processor to receive an orthographic space from the server, generate a set of candidate phoneme strings having pronunciations close to the input word in the orthographic space, and select phoneme sub-strings from the set as a pronunciation for the input word.

68. The system of claim 67, wherein the client process further causes the processor, when generating a set of candidate phoneme strings, to determine phoneme strings for the set based on closeness measure between phoneme strings for words chosen from a dictionary and contexts with in the input word.

69. The system of claim 68, wherein the client process further causes the processor to choose the words from the dictionary based on a closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary.

70. The system of claim 69, wherein the server creates the orthographic anchors by mapping words in a dictionary and a set of letter patterns that are characteristic of the words into the orthographic space.

71. The system of claim 70, wherein the server further creates a matrix of feature vectors from the letter patterns and the words, decomposes the matrix into row vectors and column vectors, and selects vectors from one of the row vectors and column vectors as the orthographic anchors.

72. The system of claim 67, wherein the client process further causes the processor, when selecting phoneme sub-strings, to align the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts and choose the phoneme sub-strings for the pronunciation based on the occurrence counts.

73. A system comprising:

a processor coupled to a memory through a bus;

a network interface coupled to the processor through the bus to communicate with a client; and

a server process executed from the memory by the processor to cause the processor to receive, from the client, an input word and a neighborhood that identifies phoneme strings having pronunciations close to the input word in an orthographic space, generate a set of candidate phoneme strings for the input word from the phoneme strings, select phoneme sub-strings from the set as a pronunciation for the input word, and return the pronunciation to the client.

74. The system of claim 73, wherein the server process further causes the processor, when generating a set of candidate phoneme strings, to determine phoneme strings for the set based on a closeness measure between the phoneme strings and contexts within the input word.

75. The system of claim 73, wherein the server process further causes the processor, when selecting phoneme sub-strings, to align the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts, and choose the phoneme sub-strings for the pronunciation based on the occurrence counts.

76. The system of claim 73, wherein the neighborhood is an orthographic neighborhood and the server process further causes the processor to create a pronunciation neighborhood from the phoneme strings for words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

77. The system of claim 76, wherein the client chooses words for the orthographic neighborhood from a dictionary based on a closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary.

78. The system of claim 77, wherein the client further creates the orthographic anchors by mapping the words in the dictionary and a set of letter patterns that are characteristic of the words into the orthographic space.

79. The system of claim 78, wherein the client further creates a matrix of feature vectors from the letter patterns and the words, decomposes the matrix into row vectors and column vectors, and selects vectors from one of the row vectors and column vectors as the orthographic anchors.

80. The system of claim 73, wherein the neighborhood is a pronunciation neighborhood created by the client from the phoneme strings for words in an orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

81. A system comprising:

a processor coupled to a memory through a bus;

an interface coupled to the processor through the bus to receive an input word;

a network interface coupled to the processor through the bus to communicate with a server; and

a client process executed from the memory by the processor to cause the processor to receive, from the server, a neighborhood that identifies phoneme strings having pronunciations close to an input word in an orthographic space, generate a set of candidate phoneme strings for the input word from the phoneme strings, select phoneme sub-strings from the set as a pronunciation for the input word, and return the pronunciation to the client.

82. The system of claim 81, wherein the client process further causes the processor, when generating a set of candidate phoneme strings, to determine phoneme strings for the set based on a closeness measure between the phoneme strings and contexts within the input word.

83. The system of claim 81, wherein the client process further causes the processor, when selecting phoneme sub-strings, to align the candidate phoneme strings on common phoneme sub-strings to produce occurrence counts, and choose the phoneme sub-strings for the pronunciation based on the occurrence counts.

84. The system of claim 81, wherein the neighborhood is an orthographic neighborhood and the client process further causes the processor to create a pronunciation neighborhood from the phoneme strings for words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

85. The system of claim 84, wherein the server chooses words for the orthographic neighborhood from a dictionary based on a closeness measure between a representation of the input word in the orthographic space and orthographic anchors corresponding to the words in the dictionary.

86. The system of claim 85, wherein the server further creates the orthographic anchors by mapping the words in the dictionary and a set of letter patterns that are characteristic of the words into the orthographic space.

87. The system of claim 86, wherein the server further creates a matrix of feature vectors from the letter patterns and the words, decomposes the matrix into row vectors and column vectors, and selects vectors from one of the row vectors and column vectors as the orthographic anchors.

88. The system of claim 81, wherein the neighborhood is a pronunciation neighborhood created by the server from the phoneme strings for words in an orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.