Unsupervised data-driven pronunciation modeling Number:7,047,193 from the United States Patent and Trademark Office (PTO) owispatent

Title: Unsupervised data-driven pronunciation modeling

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.

Patent Number: 7,047,193 Issued on 05/16/2006 to Bellegarda

---

Inventors:	Bellegarda; Jerome R. (Los Gatos, CA)
Assignee:	Apple Computer, Inc. (Cupertino, CA)
Appl. No.:	243017
Filed:	September 13, 2002

---

Current U.S. Class:	704/254
Current International Class:	G10L 15/04 (20060101)
Field of Search:	704/254

---

References Cited [Referenced By]

---

U.S. Patent Documents

	5040218		August 1991		Vitale et al.
	5268990		December 1993		Cohen et al.
	5333275		July 1994		Wheatley et al.
	5581655		December 1996		Cohen et al.
	5799276		August 1998		Komissarchik et al.
	6016471		January 2000		Kuhn et al.
	6029132		February 2000		Kuhn et al.
	6108627		August 2000		Sabourin
	6684187		January 2004		Conkie
	6778962		August 2004		Kasai et al.
	6810379		October 2004		Vermeulen et al.

Other References

Bellegarda, Jerome R., "Exploiting Latent Semantic Information in Statistical Language Modeling," Proceedings of the IEEE, Aug. 2000, pp. 1-18, no days. cited by other . Vingron, Martin, "Near-Optimal Sequence Alignment," Curr. Op. Struct. Biology, vol. 6, No. 3, Jun. 1996, 20 pp., no days. cited by other.

Primary Examiner: Abebe; Daniel
Assistant Examiner: Opsasnick; Michael N
Attorney, Agent or Firm: Blakely, Sokoloff, Taylor & Zafman LLP

---

Claims

---

What is claimed is:

1. A computerized method comprising: generating a set of candidate phoneme strings having pronunciations close to an input word in an orthographic space based on a first closeness measure between phoneme strings for words chosen from a dictionary and contexts within the input word, the choice of words 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; and selecting phoneme sub-strings from the set as a pronunciation for the input word.

2. The computerized method of claim 1 further comprising: creating the representation of the input word in the orthographic space.

3. The computerized method of claim 2, wherein the representation of the input word and the orthographic anchors are feature vectors.

4. The computerized method of claim 3, wherein the feature vector {overscore ({tilde over (v)}.sub.p for the input word {tilde over (w)}.sub.p is calculated as .times..times. ##EQU00006## where, S is a singular diagonal matrix, U is a left singular matrix, {tilde over (v)}.sub.p is a vector within a right singular matrix V.sup.T corresponding to the input word, and .sup.T denotes matrix transposition.

5. The computerized method of claim 3, wherein creating the representation comprises: mapping the input word onto the feature vector in the orthographic vector space.

6. The computerized method of claim 5, wherein mapping the input word comprises: weighing occurrences of each of a set of letter patterns within the input word.

7. The computerized method of claim 6, wherein weighing occurrences comprises: calculating a weight for the input word {tilde over (w)}.sub.p as .epsilon..times. ##EQU00007## where c.sub.i,j is the number of times letter pattern t.sub.i occurs in the input word {tilde over (w)}.sub.p, n.sub.j is the total number of letter patterns present in this word, and .epsilon..sub.i is the normalized entropy of t.sub.i in the words of the dictionary.

8. The computerized method of claim 1, wherein choosing the words comprises: filtering the orthographic anchor for each word against a pre-determined closeness threshold; and forming an orthographic neighborhood from words corresponding to orthographic anchors that meet the threshold.

9. The computerized method of claim 8, wherein choosing the words further comprises: ranking each word based on the second closeness measure.

10. The computerized method of claim 8 further comprising: creating a pronunciation neighborhood from the phoneme strings for the words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

11. The computerized method of claim 8, wherein the first closeness measure is a distance between an orthographic anchor corresponding to a word in the orthographic neighborhood and a context.

12. The computerized method of claim 1, wherein the second closeness measure comprises the cosine of the angle between the representation of the input word and an orthographic anchor.

13. The computerized method of claim 12, wherein the cosine K between the representation of the input word {overscore ({tilde over (v)}.sub.p and an orthographic anchor {overscore (v)}.sub.j is calculated using .function..times..times..times..times..times..times..times. ##EQU00008## for any 1.ltoreq.j.ltoreq.M, where M is the number of words in the dictionary, S is a singular diagonal matrix, {tilde over (v)}.sub.p is a column vector within a right singular matrix V.sup.T corresponding to the input word, v.sub.j is a column vector in V for the word corresponding to the orthographic anchor {overscore (v)}.sub.j, and .sup.T denotes matrix transposition.

14. The computerized method of claim 1, wherein the contexts are letter-based strings.

15. The computerized method of claim 14, wherein a length for the contexts and a length for the letter patterns are the same length and the first and second closeness measures are the same closeness measure.

16. The computerized method of claim 1 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.

17. The computerized method of claim 16, 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.

18. The computerized method of claim 16 further comprising: determining the letter patterns based on the words in the dictionary.

19. The computerized method of claim 1, wherein the contexts are of different lengths.

20. 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.

21. The computerized method of claim 20, wherein aligning candidate phoneme strings comprises: calculating a minimum cost alignment A(k, l) between two candidate phoneme strings .phi..sub.1 . . . .phi..sub.k . . . .phi..sub.K and .psi..sub.1 . . . .psi..sub.l . . . .psi..sub.L having a length K and L respectively.

22. The computerized method of claim 21, wherein the minimum cost alignment is calculated as A(k, l)=min {A(k-1, l-1)+C(k, l),G(i, k), H(j, l)} where C(k, l) is the cost of substituting phoneme .psi..sub.l for phoneme .phi..sub.k, g(i, k) is the cost of a gap .phi..sub.i . . . .phi..sub.k, h(j, l) is the cost of a gap .psi..sub.j . . . .psi..sub.l, .function..ltoreq..ltoreq..times..function..function. ##EQU00009## and .function..ltoreq..ltoreq..times..function..function. ##EQU00010## with initial conditions A(k, 0)=h(0, k), 1.ltoreq.k.ltoreq.K and A(0, l)=g(0, l), 1.ltoreq.l.ltoreq.L.

23. The computerized method of claim 20, wherein aligning the candidate phoneme strings produces an occurrence count for each context within the input word.

24. The computerized method of claim 23 further comprising: merging the chosen phoneme sub-strings for adjacent contexts when the phoneme sub-strings overlap.

25. A machine-readable medium having instructions to cause a machine to perform a method comprising: generating a set of candidate phoneme strings having pronunciations close to an input word in an orthographic space based on a first closeness measure between phoneme strings for words chosen from a dictionary and contexts within the input word, the choice of words 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; and selecting phoneme sub-strings from the set as a pronunciation for the input word.

26. The machine-readable medium of claim 25, wherein the method further comprises: creating the representation of the input word in the orthographic space.

27. The machine-readable medium of claim 26, wherein the representation of the input word and the orthographic anchors are feature vectors.

28. The machine-readable medium of claim 27, wherein the feature vector {overscore ({tilde over (v)}.sub.p for the input word {tilde over (w)}.sub.p is calculated as .times..times. ##EQU00011## where, S is a singular diagonal matrix, U is a left singular matrix, {tilde over (v)}.sub.p is a vector within a right singular matrix V.sup.T corresponding to the input word, and .sup.T denotes matrix transposition.

29. The machine-readable medium of claim 27, wherein creating the representation comprises: mapping the input word onto the feature vector in the orthographic vector space.

30. The machine-readable medium of claim 29, wherein mapping the input word comprises: weighing occurrences of each of a set of letter patterns within the input word.

31. The machine-readable medium of claim 30, wherein weighing occurrences comprises calculating a weight for the input word {tilde over (w)}.sub.p as .epsilon..times. ##EQU00012## where c.sub.i,j is the number of times letter pattern t.sub.i occurs in the input word {tilde over (w)}.sub.p, n.sub.j is the total number of letter patterns present in this word, and .epsilon..sub.i is the normalized entropy of t.sub.i in the words of the dictionary.

32. The machine-readable medium of claim 25, wherein choosing the words comprises: filtering the orthographic anchor for each word against a pre-determined closeness threshold; and forming an orthographic neighborhood from words corresponding to orthographic anchors that meet the threshold.

33. The machine-readable medium of claim 32, wherein choosing the words further comprises: ranking each word based on the second closeness measure.

34. The machine-readable medium of claim 32, wherein the method further comprises: creating a pronunciation neighborhood from the phoneme strings for the words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

35. The machine-readable medium of claim 32, wherein the first closeness measure is a distance between an orthographic anchor corresponding to a word in the orthographic neighborhood and a context.

36. The machine-readable medium of claim 25, wherein the second closeness measure comprises the cosine of the angle between the representation of the input word and an orthographic anchor.

37. The machine-readable medium of claim 36, wherein the cosine K between the representation of the input word {overscore ({tilde over (v)}.sub.p and an orthographic anchor {overscore (v)}.sub.j is calculated using .function..times..times..times..times..times..times..times. ##EQU00013## for any 1.ltoreq.j.ltoreq.M, where M is the number of words in the dictionary, S is a singular diagonal matrix, {tilde over (v)}.sub.p is a column vector within a right singular matrix V.sup.T corresponding to the input word, v.sub.j is a column vector in V for the word corresponding to the orthographic anchor {overscore (v)}.sub.j, and .sup.T denotes matrix transposition.

38. The machine-readable medium of claim 25, wherein the contexts are letter-based strings.

39. The machine-readable medium of claim 38, wherein a length for the contexts and a length for the letter patterns are the same length and the first and second closeness measures are the same closeness measure.

40. 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 an orthographic space to create orthographic anchors.

41. The machine-readable medium of claim 40, 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.

42. The machine-readable medium of claim 41, wherein the method further comprises: determining the letter patterns based on the words in the dictionary.

43. The machine-readable medium of claim 25, wherein the contexts are of different lengths.

44. The machine-readable medium of claim 25, 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.

45. The machine-readable medium of claim 44, wherein aligning candidate phoneme strings comprises: calculating a minimum cost alignment A(k, l) between two candidate phoneme strings .phi..sub.1 . . . .phi..sub.k . . . .phi..sub.K and .psi..sub.1 . . . .psi..sub.l . . . .psi..sub.L having a length K and L respectively.

46. The machine-readable medium of claim 45, wherein the minimum cost alignment is calculated as A(k, l)=min {A(k-1, l-1)+C(k, l), G(i, k), H(j, l)} where C(k, l) is the cost of substituting phoneme .psi..sub.l for phoneme .phi..sub.k, g(i, k) is the cost of a gap .phi..sub.i . . . .phi..sub.k, h(j, l) is the cost of a gap .psi..sub.j . . . .psi..sub.l, .function..ltoreq..ltoreq..times..function..function. ##EQU00014## and .function..ltoreq..ltoreq..times..function..function. ##EQU00015## with initial conditions A(k, 0)=h(0, k), 1.ltoreq.k.ltoreq.K and A(0, l)=g(0, l), 1.ltoreq.l.ltoreq.L.

47. The machine-readable medium of claim 44, wherein aligning the candidate phoneme strings produces an occurrence count for each context within the input word.

48. The machine-readable medium of claim 47, wherein the method further comprises: merging the chosen phoneme sub-strings for adjacent contexts when the phoneme sub-strings overlap.

49. A system comprising: a processing unit coupled to a memory through a bus; and a pronunciation modeling process executed from the memory by the processing unit to cause the processing unit to generate a set of candidate phoneme strings having pronunciations close to an input word in an orthographic space based on a first closeness measure between phoneme strings for words chosen from a dictionary and contexts within the input word, the choice of words 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, and select a common phoneme sub-string as a pronunciation for each context based on occurrences of the common sub-strings for the context.

50. The system of claim 49, wherein the pronunciation modeling process further causes the processing unit to create the representation of the input word in the orthographic space.

51. The system of claim 50, wherein the representation of the input word and the orthographic anchors are feature vectors.

52. The system of claim 51, wherein the feature vector for the input word {overscore ({tilde over (v)}.sub.p is calculated as .times..times. ##EQU00016## where, S is a singular diagonal matrix, U is a left singular matrix, {tilde over (v)}.sub.p is a vector within a right singular matrix V.sup.T correspond to the input word, and .sup.T denotes matrix transposition.

53. The system of claim 51, wherein the pronunciation modeling process further causes the processing unit, when creating the representation, to map the input word onto the feature vector in the orthographic vector space.

54. The system of claim 53, wherein the pronunciation modeling process further causes the processing unit, when mapping the input word, to weigh occurrences of each of a set of letter patterns within the input word.

55. The system of claim 54, wherein the pronunciation modeling process further causes the processing unit, when weighing occurrences, to calculate a weight for the input word {tilde over (w)}.sub.p as .epsilon..times. ##EQU00017## where c.sub.i,j is the number of times letter pattern t.sub.i occurs in the input word {tilde over (w)}.sub.p, n.sub.j is the total number of letter patterns present in this word, and .epsilon..sub.i is the normalized entropy of t.sub.i in the words of the dictionary.

56. The system of claim 49, wherein the pronunciation modeling process further causes the processing unit, when choosing the words, to filter the orthographic anchor for each word against a pre-determined closeness threshold and form an orthographic neighborhood from words corresponding to orthographic anchors that meet the threshold.

57. The system of claim 56, wherein the pronunciation modeling process further causes the processing unit, when choosing the words, to rank each word based on the second closeness measure.

58. The system of claim 56, wherein the pronunciation modeling process further causes the processing unit to create a pronunciation neighborhood from the phoneme strings for the words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

59. The system of claim 56, wherein the first closeness measure is a distance between an orthographic anchor corresponding to a word in the orthographic neighborhood and a context.

60. The system of claim 49, wherein the second closeness measure comprises the cosine of the angle between the representation of the input word and an orthographic anchor.

61. The system of claim 60, wherein the cosine K between the representation of the input word {overscore ({tilde over (v)}.sub.p and an orthographic anchor {overscore (v)}.sub.j is calculated using .function..times..times..times..times..times..times..times. ##EQU00018## for any 1.ltoreq.j.ltoreq.M, where M is the number of words in the dictionary, S is a singular diagonal matrix, {tilde over (v)}.sub.p is a column vector within a right singular matrix V.sup.T correspond to the input word, v.sub.j is a column vector in V for the word correspond to the orthographic anchor {overscore (v)}.sub.j and .sup.T denotes matrix transposition.

62. The system of claim 49, wherein the contexts in the input word are letter-based strings.

63. The system of claim 62, wherein a length for the contexts and a length for the letter patterns are the same length and the first and second closeness measures are the same closeness measure.

64. The system of claim 49, wherein the pronunciation modeling process further causes the processing unit to map 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 system of claim 64, wherein the pronunciation modeling process further causes the processing unit, when mapping the words, to create a matrix of feature vectors from the letter patterns and the words, decompose the matrix into row vectors and column vectors, and select vectors from one of the row vectors and column vectors as the orthographic anchors.

66. The system of claim 64, wherein the pronunciation modeling process further causes the processing unit to determine the letter patterns based on the words in the dictionary.

67. The system of claim 49, wherein the contexts are of different lengths.

68. The system of claim 49, wherein the pronunciation modeling process further causes the processing unit, 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.

69. The system of claim 68, wherein the pronunciation modeling process further causes the processing unit, when aligning candidate phoneme strings, to calculate a minimum cost alignment A(k, l) between two candidate phoneme strings .phi..sub.1 . . . .phi..sub.k . . . .phi..sub.K and .psi..sub.1 . . . .psi..sub.l . . . .psi..sub.L having a length K and L respectively.

70. The system of claim 69, wherein the minimum cost alignment is calculated as A(k, l)=min {A(k-1, l-1)+C(k, l)G(i, k), H(j, l)} where C(k, l) is the cost of substituting phoneme .psi..sub.l for phoneme .phi..sub.k, g(i, k) is the cost of a gap .phi..sub.i . . . .phi..sub.k, h(j, k) is the cost of a gap .psi..sub.j . . . .psi..sub.l, .function..ltoreq..ltoreq..times..function..function. ##EQU00019## and .function..ltoreq..ltoreq..times..function..function. ##EQU00020## with initial conditions A(k, 0)=h(0, k), 1.ltoreq.k.ltoreq.K and A(0, l)=g(0, l), 1.ltoreq.l.ltoreq.L.

71. The system of claim 68, wherein the pronunciation modeling process further causes the processing unit, when aligning candidate phoneme strings, to produce an occurrence count for each context within the input word.

72. The system of claim 71, wherein the pronunciation modeling process further causes the processing unit to merge the chosen phoneme sub-strings for adjacent contexts when the phoneme sub-strings overlap.

73. An apparatus comprising: means for generating a set of candidate phoneme strings having pronunciations close to an input word in an orthographic space based on a first closeness measure between phoneme strings for words chosen from a dictionary and contexts within the input word, the choice of words 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; and means for selecting phoneme sub-strings from the set as a pronunciation for the input word.

74. The apparatus of claim 73, wherein the means for generating further comprises: means for creating the representation of the input word in the orthographic space.

75. The apparatus of claim 74, wherein the representation of the input word and the orthographic anchors are feature vectors.

76. The apparatus of claim 75, wherein the means for creating the representation calculates the feature vector {overscore ({tilde over (v)}.sub.p for the input word {tilde over (w)}.sub.p as .times..times. ##EQU00021## where, S is a singular diagonal matrix, U is a left singular matrix, {tilde over (v)}.sub.p is a vector within a right singular matrix V.sup.T corresponding to the input word, and .sup.T denotes matrix transposition.

77. The apparatus of claim 75 wherein the means for creating the representation maps the input word onto the feature vector in the orthographic vector space.

78. The apparatus of claim 77, wherein the means for creating the representation further weighs occurrences of each of a set of letter patterns within the input word.

79. The apparatus of claim 78, wherein weighing occurrences comprises: calculating a weight for the input word {tilde over (w)}.sub.p as .epsilon..times. ##EQU00022## where c.sub.i,j is the number of times letter pattern t.sub.i occurs in the input word {tilde over (w)}.sub.p, n.sub.j is the total number of letter patterns present in this word, and .epsilon..sub.i is the normalized entropy of t.sub.i in the words of the dictionary.

80. The apparatus of claim 73, wherein the means for generating further comprises: means for filtering the orthographic anchor for each word against a pre-determined closeness threshold to form an orthographic neighborhood from words corresponding to orthographic anchors that meet the threshold.

81. The apparatus of claim 80, wherein the means for generating further comprises: means for ranking each word based on the second closeness measure.

82. The apparatus of claim 80, wherein the means for generating further comprises: means for creating a pronunciation neighborhood from the phoneme strings for the words in the orthographic neighborhood, the candidate phoneme strings being generated from the pronunciation neighborhood.

83. The apparatus of claim 80, wherein the first closeness measure is a distance between an orthographic anchor corresponding to a word in the orthographic neighborhood and a context.

84. The apparatus of claim 73, wherein the second closeness measure comprises the cosine of the angle between the representation of the input word and an orthographic anchor.

85. The apparatus of claim 84, wherein the cosine K between the representation of the input word {overscore ({tilde over (v)}.sub.p and an orthographic anchor {overscore (v)}.sub.j is calculated using .function..times..times..times..times..times..times..times. ##EQU00023## for any A1.ltoreq.j.ltoreq.M, where M is the number of words in the dictionary, S is a singular diagonal matrix, {tilde over (v)}.sub.p is a column vector within a right singular matrix V.sup.T corresponding to the input word, v.sub.j is a column vector in V for the word corresponding to the orthographic anchor {overscore (v)}.sub.j, and .sup.T denotes matrix transposition.

86. The apparatus of claim 73, wherein the contexts are letter-based strings.

87. The apparatus of claim 86, wherein a length for the contexts and a length for the letter patterns are the same length and the first and second closeness measures are the same closeness measure.

88. The apparatus of claim 73 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.

89. The apparatus of claim 88, wherein the means for mapping further comprises: means for creating a matrix of feature vectors from the letter patterns and the words; and means for decomposing the matrix into row vectors and column vectors, vectors from one of the row vectors and column vectors being selected as the orthographic anchors.

90. The apparatus of claim 88, further comprising: means for determining the letter patterns based on the words in the dictionary.

91. The apparatus of claim 73, wherein the contexts are of different lengths.

92. The apparatus of claim 73, wherein the means for selecting phoneme sub-strings 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.

93. The apparatus of claim 92, wherein the means for aligning candidate phoneme strings comprises: means for calculating a minimum cost alignment A(k, l) between two candidate phoneme strings .phi..sub.1 . . . .phi..sub.k . . . .phi..sub.K and .psi..sub.1 . . . .psi..sub.l . . . .psi..sub.L having a length K and L respectively.

94. The apparatus of claim 93, wherein the minimum cost alignment is calculated as A(k, l)=min {A(k-A1, l-A1)+C(k, l)G(i, k), H(j, l)} where C(k, l) is the cost of substituting phoneme .psi..sub.l for phoneme .phi..sub.k, g(i, k) is the cost of a gap .phi..sub.i . . . .phi..sub.k, is the cost of a gap .psi..sub.j . . . .psi..sub.l, .function..ltoreq..ltoreq..times..function..function. ##EQU00024## and .function..ltoreq..ltoreq..times..function..function. ##EQU00025## with initial conditions A(k, 0)=h(0, k), A1.ltoreq.k.ltoreq.K and A(0, l)=g(0, l), A1.ltoreq.l.ltoreq.L.

95. The apparatus of claim 92, wherein the means for aligning the candidate phoneme strings produces an occurrence count for each context within the input word.

96. The apparatus of claim 95 further comprising: means for merging the chosen phoneme sub-strings for adjacent contexts when the phoneme sub-strings overlap.

---

Description

---

The present application is related to U.S. patent application Ser. No. 10/242,849, filed concurrently with and assigned to the same assignee as the present application.

FIELD OF THE INVENTION

This invention relates generally to speech processing, and more particularly to pronunciation modeling.

COPYRIGHT NOTICE/PERMISSION

A portion of the disclosure of this patent document contains material which is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure as it appears in the Patent and Trademark Office patent file or records, but otherwise reserves all copyright rights whatsoever. The following notice applies to the software and data as described below and in the drawings hereto: Copyright .COPYRGT. 2002, Apple Computer, Inc., All Rights Reserved.

BACKGROUND OF THE INVENTION

Pronunciation modeling is the process of assigning to each word in a given vocabulary a suitable sequence of phonemes (or phones). Good pronunciation modeling is critical to all speech-to-text and text-to-speech applications. In automatic speech recognition (ASR), the search module relies on phoneme sequences stored in the underlying (phonetic) dictionary to select appropriate acoustic models against which to score the input utterance. Similarly, in text-to-speech (TTS) synthesis, phonemic expansion is required for the selection of the proper TTS units from which to generate the desired waveform.

Given an input vocabulary, there are two complementary ways to construct a set of pronunciations. The most obvious method is to have one or several trained linguists manually create each entry. This is typically a time-consuming task, often prone to inconsistencies, and inherently dependent on the language considered. Yet, over the past decade, phonetic dictionaries have been in increasingly high demand as many spoken language applications reached large-scale deployment worldwide. As a result, medium (10,000) to large (100,000) dictionaries, of varying quality, are now available for most major languages.

The second method is to automatically derive pronunciations from the word orthography, e.g., the sequence of letters used to convey the word. This is necessary for the real-time processing of out-of-vocabulary words, for which no entry exists in the underlying dictionary. In fact, the dynamic nature of language makes it an instrumental part of any system. Automatic pronunciation modeling relies on a set of so-called letter-to-sound rules, whose role is to capture language regularity and properly encapsulate it within a small number of general principles. Letter-to-sound rules come in two basic flavors. They can be (hand-)written by trained linguists, as in the case of morphological decomposition for example; this approach, however, tends to suffer from the same drawbacks as above. Or they can be primarily data-driven, whereby a statistical algorithm leverages an existing dictionary to model the salient relationships between orthography and pronunciation.

In the latter case, the state-of-the-art is to train a decision tree to classify each letter sequence into the appropriate phoneme sequence. During training, the decision tree is presented with sequence pairs, aligned with the help of (language-dependent) letter mappings. During classification, the tree is traversed on the basis of questions asked about the context of each letter, until a leaf corresponding to a particular phoneme (or phoneme string) is reached. The final pronunciation is simply the concatenation of the results obtained at each leaf. Although data-driven, this procedure is not really unsupervised, since the letter mappings rely on expert human knowledge of the language considered.

"Letter-to-sound" rules work reasonably well for words that are fairly close to exemplars seen in training, but they often break down otherwise. The primary reason why current "letter-to-sound" rule implementations generalize poorly is that they attempt to capture language regularity and encapsulate it within a small number of general principles. This can be viewed as a "top-down" approach. The immediate consequence is that all rare phenomena are presumed "irregular" and, accordingly, ignored. This is a major drawback in a situation where most occurrences are infrequent, as in name pronunciation.

Perhaps even more importantly, by construction decision trees only ask contextual questions associated with phenomena that are sufficiently well represented in the training data. Contexts rarely seen in the underlying dictionary tend to be overlooked, regardless of their degree of similarity or relevance. For out-of-vocabulary words that largely conform to the general patterns of the language, as observed in the training words, this is relatively inconsequential. But many other words, such as names (especially those of foreign origin), may comprise a number of patterns rarely seen in the dictionary, for which this limitation may be more deleterious. To illustrate, consider the name of Indian origin "Krishnamoorthy," whose correct pronunciation is given by: krIHSnAXm1UHrTIY (1) using a standard phonetic notation known as AppleBet from Apple Computer, Inc., the assignee of the present invention. In contrast, the pronunciation produced by a typical letter-to-sound rule decision tree trained on 56,000 names of predominantly Western European origin and which produces average results, is given by: krIHSnAEm1UWUXrDIY (2) In particular three errors stand out in (2): the schwa "AX" in sixth position is replaced by the full vowel "AE," the penultimate unvoiced "T" is replaced by the voiced version "D," and the stressed vowel "1UH" is replaced by the improper compound "1UW UX." These errors can all be traced to the poor generalization properties of the decision tree framework. Specifically, the ending "UX r D IY" results from the influence of a number of names in the training dictionary ending in "orthy," such as "Foxworthy." The vowel compound comes from the inability of this pattern to account for "oo," hence the awkward attempt to have it both ways by concatenating the two vowels. Finally, the full vowel "AE," commonly seen in names like "McNamara," points to an obvious failure to connect "Krishnamoorthy" with the more semantically related "Krishna." This example underscores the importance of exploiting all relevant contexts, regardless of how sparsely seen they may have been in the training data, to increase the ability of a pronunciation model to generalize.

SUMMARY OF THE INVENTION

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. Thus, the pronunciation modeling recognizes global phoneme patterns, instead of relying on the sequence of the letters in the words as in the letter-to-sound rule decision trees, and leverages relevant contexts, dense and sparse, within the underlying dictionary to enforce local consistency with the contexts of the out-of-vocabulary words.

The present invention is described in conjunction with systems, clients, servers, methods, and machine-readable media of varying scope. In addition to the aspects of the present invention described in this summary, further aspects of the invention will become apparent by reference to the drawings and by reading the detailed description that follows.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a diagram illustrating a system-level overview of an embodiment of the invention;

FIGS. 2A E are flowcharts of method to be performed by a machine according to the embodiment of the invention of FIG. 1;

FIGS. 3A D are diagrams of matrices produced and used by the methods of FIGS. 2A E;

FIG. 4A is a diagram of one embodiment of an operating environment suitable for practicing the present invention; and

FIG. 4B is a diagram of one embodiment of a computer system suitable for use in the operating environment of FIG. 4A.

DETAILED DESCRIPTION OF THE INVENTION

In the following detailed description of embodiments of the invention, reference is made to the accompanying drawings in which like references indicate similar elements, and in which is shown by way of illustration specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention, and it is to be understood that other embodiments may be utilized and that logical, mechanical, electrical, functional, and other changes may be made without departing from the scope of the present invention. The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the present invention is defined only by the appended claims.

A system level overview of the operation of an embodiment of the invention is described by reference to FIG. 1. A pronunciation modeling system 100 utilizes "orthographic neighborhoods" to generate a pronunciation of an out-of-vocabulary word through a bottom-up approach. For a given out-of-vocabulary word, its orthographic neighborhood is defined as the set of in-vocabulary words that are sufficiently "close" to the out-of-vocabulary word as measured by a suitable metric. In the embodiments described herein, the metric is adopted from latent semantic analysis (LSA) but any algorithm that provides a global analysis of the orthography of words is applicable.

In the embodiment illustrated in FIG. 1, the pronunciation modeling system 100 comprises two components: a training component 103 and a pronunciation component 105. When the pronunciation modeling system 100 is being trained, a training module 109 uses LSA to determine what letter strings are most characteristic of words, and to map all training words in a dictionary 107 onto all characteristic letter strings. The resulting vector space 111 contains a plurality of orthographic anchors, one for each "in-vocabulary" word, as determined automatically from the underlying dictionary 107. The orthographic anchors may be viewed as global letter patterns without letter sequence information, e.g., the sequence "cat" and the sequence "tac" will map to the same global pattern. The vector space 111 is referred to herein as "orthographic" instead of "semantic," as is commonly used with LSA, because the vector space represents orthographic information for the in-vocabulary words. It will be appreciated that the dictionary 107 contains both the in-vocabulary words and their associated phonetic expansions in the form of phoneme strings.

During operation of the pronunciation modeling system 100, an out-of-vocabulary (OOV) word 113 is mapped 115 into the orthographic vector space 111 using LSA. The representation of the out-of-vocabulary word 113 in the orthographic vector space 117 is compared to each orthographic anchor, and the corresponding "closeness" evaluated by a similarity measuring module 119. Assuming the closeness is high enough, the associated in-vocabulary word is added to the orthographic neighborhood 121 of the out-of-vocabulary word.

Once the orthographic neighborhood 121 is populated, a corresponding set of phoneme strings from the dictionary 107 is assembled, illustrated as pronunciation neighborhood 123. The phoneme strings for the in-vocabulary words in the pronunciation neighborhood 123 have the property to contain at least one sub-string which is "locally close" to the pronunciation sought. Thus, the pronunciation can be assembled from the phoneme sub-strings in the pronunciation neighborhood. The more frequently a particular sub-string appears in a particular position within the out-of-vocabulary word, the more likely it is to belong to the correct phonetic expansion ("maximum likelihood estimate"). Because sequence information is not preserved in the orthographic neighborhood 121, the aligning module 125 aligns the phoneme strings along their common sub-strings. The sub-string with the maximum likelihood estimate at each position is assembled into the pronunciation 127 of the out-of-vocabulary word 113.

Thus, the use of orthographic neighborhoods in the pronunciation modeling system 100 takes advantage of orthographic information of the in-vocabulary words without requiring explicit language-dependent letter mappings. The pronunciation modeling system recognizes global phoneme patterns across the in-vocabulary and out-of-vocabulary words, instead of relying on the sequence of the letters in the words. The alignment of phoneme strings that produces the pronunciation is between phoneme strings instead of between a phoneme string and a letter string. Furthermore, the pronunciation modeling system 100 is data-driven and requires little or no human supervision. While the invention is not limited to any particular arrangement of components, for sake of clarity a simplified pronunciation modeling system having separate training and pronunciation components has been described.

Methods to be performed by the embodiment of the pronunciation modeling system 100 illustrated in FIG. 1 are described with reference to flowcharts shown in FIGS. 2A E. FIG. 2A illustrates the acts to be performed by a machine, such as a computer, in executing an embodiment of an orthographic space method 200 to create the orthographic anchors for the in-vocabulary words in an orthographic vector space. FIG. 2B illustrates the acts to be performed in executing an embodiment of a pronunciation model 210 to create a pronunciation for an out-of-vocabulary word, with FIGS. 2C E illustrating the processing of FIG. 2B in more detail.

Referring first to FIG. 2A, the method 200 that creates the orthographic anchors in the orthographic vector space is described. Assuming a dictionary containing a vocabulary comprising the M most frequent words in the language, the method 200 creates a set of all strings of n letters that can be produced from this vocabulary (including markers for word beginning and ending) at block 201, with N being the number of strings in the set. Typically, M varies between 10,000 and 100,000; in one embodiment in which n=3, N is on the order of 10,000.

At block 203, the method 200 constructs a co-occurrence matrix N.times.M matrix W 300 of feature vectors as shown in FIG. 3A, in which the entries w.sub.ij suitably reflect the extent to which each n-letter string t.sub.j in the set appeared in each word w.sub.j in the vocabulary. Each instance of a letter string in a word is weighted because two n-letter strings appearing with the same count in a particular word do not necessarily convey the same amount of information. In one embodiment, w.sub.ij is defined as

.times. ##EQU00001## where c.sub.i,j is the number of times t.sub.i occurs in word w.sub.j, n.sub.j is the total number of n-letter strings present in this word, and .epsilon..sub.i is the normalized entropy of t.sub.i in the vocabulary. The global weighting 1-.epsilon..sub.i is subordinated to the distribution of the n-letter strings in the entire vocabulary. It will be appreciated that the invention is not limited to this particular global weighting and other weighting algorithms may be used, such as a logarithm transformation, or any monotonously increasing function. Additionally, embodiments that omit the weighting are contemplated as within the scope of the invention.

At block 205 the method 200 decomposes the co-occurrence matrix W 300 into left and right singular matrices, illustrated in FIG. 3B, using singular value decomposition (SVD) W=USV.sup.T (4) where U is the (N.times.R) left singular matrix 301 with row vectors u.sub.i (1.ltoreq.i.ltoreq.N), S is the (R.times.R) diagonal matrix 303 of singular values s.sub.1.gtoreq.s.sub.2.gtoreq. . . . .gtoreq.S.sub.R>0, V.sup.T is the (M.times.R) right singular matrix 305 with column vectors v.sub.j(1.ltoreq.j.ltoreq.M), R<<M, N is the order of the decomposition, and .sup.T denotes matrix transposition. The dimension R is bounded from above by the rank of the co-occurrence matrix W 300, and from below by the amount of distortion tolerable in the decomposition. In one embodiment, R is in the range of 50 to 100.

After appropriate scaling of the singular matrices 301, 305 by the singular values of the diagonal matrix 303 at block 207, the rank-R decomposition defines a mapping between (i) the set of n-letter strings and the R-dimensional vectors .sub.i=u.sub.iS (1.ltoreq.i.ltoreq.N), and (ii) the set of words in the vocabulary and the R-dimensional vectors {overscore (v)}.sub.j=v.sub.jS (1.ltoreq.j.ltoreq.M). Thus, the R-dimensional orthographic vector space spanned by the vectors .sub.l and {overscore (v)}.sub.j minimally describes the underlying vocabulary and the set of n-letter strings. The R-dimensional vectors {overscore (v)}.sub.j are used as the orthographic anchors for the in-vocabulary words (block 209). It will be appreciated that if the co-occurrence matrix W 300 were transposed, i.e., the matrix would be M.times.N, the R-dimensional vectors .sub.l would be used as the orthographic anchors.

Turning now to FIG. 2B, the pronunciation model method 210 uses the orthographic vector space 111 to generate a set of candidate phoneme strings having pronunciations close to the out-of-vocabulary word (block 211). The pronunciation for the out-of-vocabulary word is assembled from selected phoneme sub-strings in the set (block 213). In one embodiment, the processing represented by block 211 is illustrated in FIGS. 2C D and in FIG. 2E at blocks 251 until 255, and the processing represented by block 213 is illustrated in FIG. 2E at blocks 257 until 265, but the invention is not limited by the particular arrangement of processing blocks shown therein. Furthermore, the processing represented in FIGS. 2C D is described in conjunction with data structures illustrated in FIGS. 3C D but the invention is not limited to any particular data structure.

One embodiment of a mapping method 220 that creates the representation of a given out-of-vocabulary word in the orthographic vector space 111 is illustrated in FIG. 2C. Assuming an out-of-vocabulary word {tilde over (w)}.sub.p where p>M, weighted counts for the out-of-vocabulary word are computed as in block 201 in FIG. 2A with j=p (block 221). The resulting feature vector, a column vector of dimension N, can be thought of as an additional column of the matrix W 300 as illustrated in FIG. 3C. Assuming the matrices U 301 and S 303 do not change, the SVD expansion (4) implies

.times. ##EQU00002## where the R-dimensional vector {tilde over (v)}.sub.p.sup.T acts as an additional column of the matrix V.sup.T as illustrated in FIG. 3D. This in turn leads to the definition of the vector

.times..times. ##EQU00003## for any out-of-vocabulary word {tilde over (w)}.sub.p in the orthographic vector space (block 223). As previously described in conjunction with FIG. 2A, the weighing at block 221 may use a different algorithm or omitted altogether.

Referring now to FIG. 2D, one embodiment of a neighborhood method 230 that creates the orthographic neighborhood, and the corresponding pronunciation neighborhood, for the out-of-vocabulary word is described. The relative positions of the orthographic anchors in the orthographic vector space reflect a parsimonious encoding of the orthography used in the training data as described above in conjunction with FIG. 2A. Thus, any out-of-vocabulary word represented by a vector "close" (in some suitable metric) to a particular orthographic anchor in the orthographic vector space can be expected to be related to the corresponding in-vocabulary word, and conversely any in-vocabulary word whose orthographic anchor is "close" to the particular vector representing the out-of-vocabulary word in the space will tend to be associated with the corresponding out-of-vocabulary word.

Accordingly, at block 231, the method 230 determines a measure of closeness for each orthographic anchor with respect to the representation of the out-of-vocabulary word. In one embodiment, the closeness measure is calculated as the cosine of the angle between the representation of the out-of-vocabulary word {overscore ({tilde over (v)}.sub.p and each orthographic anchor {overscore (v)}.sub.j

.function..times..times..times..times..times..times..times. ##EQU00004## for any 1.ltoreq.j.ltoreq.M. However, the invention is not so limited and any closeness measure, such as Euclidean and Gaussian (also known as Malahanobis), based on "bona fide" distances defined on a vector space may be used. A "bona fide" distance is one that conforms to a set of three mathematical properties as is well-known in the art.

The orthographic anchors are ranked in decreasing order of their closeness at block 233. The closeness measure of each orthographic anchor (block 235) is evaluated against pre-determined closeness threshold (block 237). The in-vocabulary word corresponding to each orthographic anchor that meets the closeness threshold is added to the orthographic neighborhood (block 239). Once the appropriately close orthographic anchors have been evaluated (block 241), the associated pronunciation neighborhood for the given out-of-vocabulary word is created from the corresponding phoneme strings for the words in the orthographic neighborhood using the dictionary (block 243). One of skill in the art will immediately recognize that a word in the vocabulary may have more than one pronunciation and thus may be represented in the pronunciations neighborhood by multiple phoneme strings. Furthermore, it will be that appreciated because the orthographic anchors are ranked at block 233, the loop represented by blocks 235 until 241 may be terminated as soon as one of the anchors fails to meet the closeness threshold. Additionally, one of skill in the art will immediately understand that the orthographic neighborhood may be more or less populated depending on the filtering criteria applied to the orthographic anchors and on the number of orthographic anchors in the orthographic space.

Turning now to FIG. 2E, one embodiment of an alignment method 250 that arranges phoneme sub-strings from the pronunciation neighborhood to produce a pronunciation for the out-of-vocabulary word is described. By construction, each phoneme string in the pronunciation neighborhood contains at least one sub-string that is germane to the out-of-vocabulary word considered. The method 250 exploits contexts within the word to align the phoneme strings in the pronunciation neighborhood to find the germane sub-strings. As an aid to understanding the following description, assume the contexts are letter-based, e.g., a particular letter and the immediately preceding and succeeding letter or marker, but the invention is not so limited.

The alignment method 250 starts at the beginning of the out-of-vocabulary word and moves left-to-right through the contexts (block 251), selecting for each context, a set of candidate phoneme strings in the pronunciation neighborhood that are close, in some measure, to the context (block 253). In one embodiment, the context is the same length n as the letter strings mapped into the orthographic vector space and the candidate phoneme strings are selected using the closeness measure described above in conjunction with FIG. 2D. Alternate embodiments that use other closeness measurements for strings, such as Hamming and Levenshtein (also know as "edit") distances, are also contemplated as within the scope of the invention. The first set of candidate phoneme strings correspond to orthographic anchors containing a beginning-of-word marker and the last set of candidate phoneme strings correspond to orthographic anchors containing an end-of-word marker (as described above in conjunction with FIG. 2A). It will be appreciated that the appropriate phoneme strings in the pronunciation neighborhood may be tagged with beginning and end word markers.

Once the set of candidate phoneme strings have been selected for each context (block 255), the method 250 evaluates each set of candidate phoneme strings (block 257) by aligning them on common sub-strings (block 259) and determining the most common sub-string for each set (block 261). The most common sub-string for each set will be used as the pronunciation for that context of the out-of-vocabulary word. Additionally, the most common sub-strings for adjacent contexts are merged when there is overlap between them at block 265, using an algorithm such as "dynamic time warping" as is well-known in the art. It will be appreciated that the merge processing represented by block 265 may be omitted. Furthermore, although the merge processing is illustrated as occurring after all the sets have been evaluated (block 263), one of skill in the art will immediately recognize that it may be performed as part of the set evaluation loop.

In one embodiment of the processing at block 263, the method 250 uses a variation of a sequence analysis technique originally applied in the field of bio-informatics to align specific regions of two protein sequences, with provisions for dealing with gaps in the alignment. Assume, without loss of generality, the need to align two phoneme strings .PHI..sub.1 . . . .PHI..sub.k . . . .PHI..sub.K and .psi..sub.1 . . . .psi..sub.l . . . .psi..sub.L (of length K and L, respectively) from the pronunciation neighborhood, and de-note by A(k, l) the minimum cost alignment between .PHI..sub.1 . . . .PHI..sub.k . . . .PHI..sub.K and .psi..sub.1 . . . .psi..sub.l . . . .psi..sub.L. If C(k, l) is the cost of substituting phoneme .psi..sub.l for phoneme .PHI..sub.k, g(i, k) the cost of a gap .PHI..sub.i . . . .PHI..sub.k in the first string, and h(j, l) the cost of a gap .psi..sub.j . . . .psi..sub.l in the second string, a recursive alignment equation can be written as

.function..times..function..function..function..function..times..times..fu- nction..ltoreq..ltoreq..times..function..function..function..ltoreq..ltore- q..times..function..function. ##EQU00005## with initial conditions A(k, 0)=h(0, k), 1.ltoreq.k.ltoreq.K and A(0, l)=g(0, l), 1.ltoreq.l.ltoreq.L.

In an alternate embodiment, the processing at blocks 251 until 255 is repeated using different granularities of contexts within the out-of-vocabulary word. The candidate phoneme strings for the contexts at a particular level of granularity form a "mini-pronunciation neighborhood" for that level, which may be arranged in a hierarchy. The phoneme strings for each context in each mini-neighborhood are aligned to find the common sub-strings (blocks 257 until 263). The resulting pronunciation could contain sub-strings from different mini-pronunciation neighborhoods.

It will be appreciated that the alignment method 250 may proceed in a pattern other than left-to-right through the input word, e.g., from both ends to the middle or from the middle outward, using the beginning-of-word and end-of word markers, and such embodiments are contemplated as within the scope of the invention.

In practice, the methods described above may constitute one or more programs made up of machine-executable instructions. Describing the methods with reference to the flowcharts in FIGS. 2A D enables one skilled in the art to develop such programs, including such instructions to carry out the operations (acts) represented by the logical blocks on suitably configured machines (the processor of the machine executing the instructions from machine-readable media). The machine-executable instructions may be written in a computer programming language or may be embodied in firmware logic or in hardware circuitry. If written in a programming language conforming to a recognized standard, such instructions can be executed on a variety of hardware platforms and for interface to a variety of operating systems. In addition, the present invention is not described with reference to any particular programming language. It will be appreciated that a variety of programming languages may be used to implement the teachings of the invention as described herein. Furthermore, it is common in the art to speak of software, in one form or another (e.g., program, procedure, process, application, module, logic . . . ), as taking an action or causing a result. Such expressions are merely a shorthand way of saying that execution of the software by a machine causes the processor of the machine to perform an action or produce a result. It will be further appreciated that more or fewer processes may be incorporated into the methods illustrated in FIGS. 2A D without departing from the scope of the invention and that no particular order is implied by the arrangement of blocks shown and described herein.

In particular, the methods of FIGS. 2A D may be modules that are executed from a memory within a computer system such as illustrated in FIG. 4B. The logical representation of the matrices illustrated in FIGS. 3A D, and the feature vectors used as the orthographic anchors and the representation of the out-of-vocabulary word by the methods may also be resident in the memory. It will be appreciated that the orthographic anchors in the orthographic vector space created during the training process may be more permanently stored on a mass storage device and retrieved into memory when required to evaluate an out-of-vocabulary word.

The following description of FIGS. 4A B is intended to provide an overview of computer hardware and other operating components suitable for performing the methods of the invention described above, but is not intended to limit the applicable environments. One of skill in the art will immediately appreciate that the invention can be practiced with other computer system configurations, including hand-held devices, multiprocessor systems, microproces