Unsupervised data-driven pronunciation modeling
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.
1. A computerized method comprising:
receiving pronunciation data for an out-of-vocabulary word, the pronunciation data comprising phoneme sub-strings selected from candidate phoneme strings having pronunciation data associated with orthographic anchors that are close to a vector representation of the out-of-vocabulary word in an orthographical vector space defined by a dictionary; and
reproducing the pronunciation data for the out-of-vocabulary word as an audible signal.
2. The computerized method of claim 1 further comprising:
storing the pronunciation data for subsequent reproduction.
3. The computerized method of claim 1 , wherein the orthographic vector space comprises the vector representation of the out-of-vocabulary word and the orthographic anchors corresponding to in-vocabulary words.
4. The computerized method of claim 3 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
5. The computerized method of claim 4 , wherein the feature vector {tilde over (v)} p for the out-of-vocabulary word {tilde over (w)} p is calculated as
{tilde over (v)} p ={tilde over (v)} p S={tilde over (w)} p T U
where, S is a singular diagonal matrix, U is a left singular matrix, {tilde over (v)} p is a vector within a right singular matrix V T corresponding to the out-of-vocabulary word, and T denotes matrix transposition.
6. The computerized method of claim 1 , wherein the dictionary comprises phoneme strings for in-vocabulary words.
7. A computerized method comprising:
receiving an orthographical vector space comprising a vector representation of an out-of-vocabulary word and orthographic anchors for in-vocabulary words, the orthographic vector space defined by a dictionary;
selecting phoneme sub-strings from the dictionary according to a closeness measure between the vector representation of the out-of-vocabulary word and the orthographic anchors for the in-vocabulary words in the orthographic vector space; and
generating a pronunciation data for the out-of-vocabulary word from the selected phoneme sub-strings.
8. The computerized method of claim 7 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
9. The computerized method of claim 8 , wherein the feature vector {tilde over (v)} p for the out-of-vocabulary word {tilde over (w)} p is calculated as
{tilde over (v)} p ={tilde over (v)} p S={tilde over (w)} p T U
where, S is a singular diagonal matrix, U is a left singular matrix, {tilde over (v)} p is a vector within a right singular matrix V T corresponding to the out-of-vocabulary word, and T denotes matrix transposition.
10. The computerized method of claim 7 , wherein selecting phoneme sub-strings comprises:
forming an orthographic neighborhood from in-vocabulary words corresponding to orthographic anchors that satisfy the closeness measure; and
creating a pronunciation data neighborhood from phoneme strings for the in-vocabulary words in the orthographic neighborhood, the phoneme strings having at least one phoneme sub-string in the dictionary.
11. The computerized method of claim 10 , wherein selecting phoneme sub-strings further comprises:
selecting phoneme strings in the pronunciation data neighborhood for each context within the out-of-vocabulary word;
aligning the selected phoneme strings on common phoneme sub-strings; and
selecting a phoneme sub-string for each context from the common phoneme sub-strings.
12. The computerized method of claim 11 further comprising:
merging the phoneme sub-strings for adjacent contexts when two phoneme sub-strings overlap.
13. The computerized method of claim 11 , wherein aligning the phoneme strings comprises:
calculating a minimum cost alignment A(k, l) between two phoneme strings φ 1 . . . φ k . . . φ K and ψ 1 . . . ψ l . . . ψ L having a length K and L respectively.
14. The computerized method of claim 13 , 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 ψ l for phoneme ψ k , g(i, k) is the cost of a gap φ i . . . φ k , h(j, l) is the cost of a gap ψ j . . . ψ i ,
G
(
i
,
k
)
=
min
0
≤
i
≤
k
-
1
{
A
(
i
,
l
)
+
g
(
i
,
k
)
}
,
and
H
(
j
,
l
)
=
min
0
≤
j
≤
l
-
1
{
A
(
k
,
j
)
+
h
(
j
,
l
)
}
,
with initial conditions A(k, 0)=h(0, k), 1≦k≦K and A(0, l)=g(0, l), 1≦l≦L.
15. The computerized method of claim 7 , wherein the closeness measure comprises the cosine of the angle between the representation of the out-of-vocabulary word and an orthographic anchor.
16. The computerized method of claim 15 , wherein the cosine K between the representation of the out-of-vocabulary word {tilde over (v)} p and an orthographic anchor v j is calculated using
K
=
(
v
_
~
p
,
v
_
j
)
=
cos
(
v
~
p
S
,
v
j
S
)
=
v
~
p
S
2
v
j
T
v
~
p
S
v
j
S
for any 1≦j≦M, where M is the number of in-vocabulary words in the dictionary, S is a singular diagonal matrix, {tilde over (v)} p is a column vector within a right singular matrix V T corresponding to the input word, v j is a column vector in V for the word corresponding to the orthographic anchor v j , and T denotes matrix transposition.
17. A computerized method comprising:
storing pronunciation data for an out-of-vocabulary word, the pronunciation data comprising phoneme sub-strings selected from candidate phoneme strings having pronunciation data associated with orthographic anchors that are close to a vector representation of the out-of-vocabulary word in an orthographical vector space defined by a dictionary; and
transmitting the pronunciation data for the out-of-vocabulary word as an audible signal.
18. The computerized method of claim 17 , wherein the vector orthographic space comprises the vector representation of the out-of-vocabulary word and the orthographic anchors corresponding to in-vocabulary words.
19. The computerized method of claim 18 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
20. The computerized method of claim 17 , wherein the dictionary comprises phoneme strings for in-vocabulary words.
21. A computer-readable storage medium storing computer-executable instructions which, when executed, cause a processing system to perform a method comprising:
receiving pronunciation data for an out-of-vocabulary word, the pronunciation data comprising phoneme sub-strings selected from candidate phoneme strings have pronunciation data associated with orthographic anchors that are close to a vector representation of the out-of-vocabulary word in an orthographical vector space defined by a dictionary; and
reproducing the pronunciation data for the out-of-vocabulary word as an audible signal.
22. The computer-readable storage medium of claim 21 , wherein the method further comprises:
storing the pronunciation data for subsequent reproduction.
23. The computer-readable storage medium of claim 21 , wherein the orthographic vector space comprises the vector representation of the out-of-vocabulary word and orthographic anchors correspond to in-vocabulary words.
24. The computer-readable storage medium of claim 23 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
25. The computer-readable storage medium of claim 21 , wherein the dictionary comprises phoneme strings for in-vocabulary words.
26. A computer-readable storage medium storing computer-executable instructions comprising:
an instruction to receive pronunciation data for an out-of-vocabulary word, the pronunciation data comprising phoneme sub-strings selected from candidate phoneme strings have pronunciation data associated with orthographic anchors that are close to a vector representation of the out-of-vocabulary word in an orthographical vector space defined by a dictionary; and
an instruction to reproduce the pronunciation data for the out-of-vocabulary word as an audible signal.
27. The computer-readable storage medium of claim 26 further comprising:
an instruction to store the pronunciation data for subsequent reproduction.
28. The computer-readable storage medium of claim 26 , wherein the orthographic vector space comprises the vector representation of the out-of-vocabulary word and the orthographic anchors corresponding to in-vocabulary words.
29. The computer-readable storage medium of claim 28 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
30. The computer-readable storage medium of claim 26 , wherein the dictionary comprises phoneme strings for in-vocabulary words.
31. A computer-readable storage medium storing computer-executable instructions which, when executed, cause a processing system to perform a method comprising:
receiving an orthographical vector space comprising a vector representation of an out-of-vocabulary word and orthographic anchors for in-vocabulary words, the orthographic vector space defined by a dictionary;
selecting phoneme sub-strings from the dictionary accord to a closeness measure between the vector representation of the out-of-vocabulary word and the orthographic anchors for the in-vocabulary words in the orthographic vector space; and
generating pronunciation data for the out-of-vocabulary word from the selected phoneme sub-strings.
32. The computer-readable storage medium of claim 31 , wherein the representation of the out-of vocabulary word and the orthographic anchors are feature vectors.
33. The computer-readable storage medium of claim 31 , wherein selecting phoneme sub-strings comprises:
forming an orthographic neighborhood from in-vocabulary words correspond to orthographic anchors that satisfy the closeness measure; and
creating a pronunciation data neighborhood from phoneme strings for the in-vocabulary words in the orthographic neighborhood, the phoneme strings have at least one phoneme sub-string in the dictionary.
34. The computer-readable storage medium of claim 33 , wherein selecting phoneme sub-strings comprises:
selecting phoneme strings in the pronunciation data neighborhood for each context within the out-of-vocabulary word;
aligning the selected phoneme strings on common phoneme sub-strings; and
selecting a phoneme sub-string for each context from the common phoneme sub-strings.
35. The computer-readable storage medium of claim 34 , wherein the method further comprises:
merging the phoneme sub-strings for adjacent contexts when two phoneme sub-strings overlap.
36. A computer-readable storage medium storing computer-executable instructions comprising:
an instruction to receive an orthographical vector space comprising a vector representation of an out-of vocabulary word and orthographic anchors for in-vocabulary words, the orthographic vector space defined by a dictionary;
an instruction to select phoneme sub-strings from the dictionary accord to a closeness measure between the vector representation of the out-of-vocabulary word and the orthographic anchors for the in-vocabulary words; and
an instruction to generate a pronunciation data for the out-of vocabulary word from the selected phoneme sub-strings.
37. The computer-readable storage medium of claim 36 , wherein the representation of the out-of vocabulary word and the orthographic anchors are feature vectors.
38. The computer-readable storage medium of claim 36 , wherein the instruction to select phoneme sub-strings comprises:
an instruction to form an orthographic neighborhood from in-vocabulary words correspond to orthographic anchors that satisfy the closeness measure; and
an instruction to create a pronunciation data neighborhood from phoneme strings for the in-vocabulary words in the orthographic neighborhood, the phoneme strings have at least one phoneme sub-string in the dictionary.
39. The computer-readable storage medium of claim 38 , wherein the instruction to select phoneme sub-strings further comprises:
an instruction to select phoneme strings in the pronunciation data neighborhood for each context within the out-of vocabulary word;
an instruction to align the selected phoneme strings on common phoneme sub-strings; and
an instruction to select a phoneme sub-string for each context from the common phoneme sub-strings.
40. The computer-readable storage medium of claim 39 further comprising:
an instructions to merge the phoneme sub-strings for adjacent contexts when two phoneme sub-strings overlap.
41. A computer-readable storage medium storing computer-executable instructions which, when executed, cause a data processing system to perform a method comprising:
storing pronunciation data for an out-of-vocabulary word, the pronunciation data comprising phoneme sub-strings selected from candidate phoneme strings having pronunciation data associated with orthographic anchors that are close to a vector representation of the out-of-vocabulary word in an orthographical space defined by a dictionary; and
transmitting the pronunciation data for the out-of-vocabulary word as an audible signal.
42. The computer-readable storage medium of claim 41 , wherein the orthographic vector space comprises the vector representation of the out-of-vocabulary word and orthographic anchors corresponding to in-vocabulary words.
43. The computer-readable storage medium of claim 42 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
44. The computer-readable storage medium of claim 41 , wherein the dictionary comprises phoneme strings for in-vocabulary words.
45. A computer-readable storage medium storing computer-executable instructions comprising:
an instruction to store pronunciation data for an out-of-vocabulary word, the pronunciation data comprising phoneme sub-strings selected from candidate phoneme strings having pronunciation data associated with orthographic anchors that are close to a vector representation of the out-of-vocabulary word in an orthographical vector space defined by a dictionary; and
an instruction to transmit the pronunciation data for the out-of-vocabulary word as an audible signal.
46. The computer-readable storage medium of claim 45 , wherein the orthographic vector space comprises the representation of the out-of-vocabulary word and the orthographic anchors corresponding to in-vocabulary words.
47. The computer-readable storage medium of claim 46 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
48. The computer-readable storage medium of claim 45 , wherein the dictionary comprises phoneme strings for in-vocabulary words.
49. An apparatus comprising:
means for receiving pronunciation data for an out-of-vocabulary word, the pronunciation data comprising phoneme sub-strings selected from candidate phoneme strings having pronunciation data associated with orthographic anchors that are close to a vector representation of the out-of-vocabulary word in an orthographical vector space defined by a dictionary; and
means for reproducing the pronunciation data for the out-of-vocabulary word as an audible signal.
50. The apparatus of claim 49 further comprising:
means for storing the pronunciation data for subsequent reproduction.
51. The apparatus of claim 49 , wherein the orthographic vector space comprises the vector representation of the out-of-vocabulary word and the orthographic anchors corresponding to in-vocabulary words.
52. The apparatus of claim 51 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
53. The apparatus of claim 49 , wherein the dictionary comprises phoneme strings for in-vocabulary words.
54. An apparatus comprising:
means for receiving an orthographical vector space comprising a vector representation of an out-of-vocabulary word and orthographic anchors for in-vocabulary words, the orthographic vector space defined by a dictionary;
means for selecting phoneme sub-strings from the dictionary according to a closeness measure between the vector representation of the out-of-vocabulary word and the orthographic anchors for the in-vocabulary words in the orthographical vector space; and
means for generating a pronunciation data for the out-of-vocabulary word from the selected phoneme sub-strings.
55. The apparatus of claim 54 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
56. The apparatus of claim 54 , wherein the means for selecting phoneme sub-strings comprises:
means for forming an orthographic neighborhood from in-vocabulary words corresponding to orthographic anchors that satisfy the closeness measure; and
means for creating a pronunciation data neighborhood from phoneme strings for the in-vocabulary words in the orthographic neighborhood, the phoneme strings having at least one phoneme sub-string in the dictionary.
57. The apparatus of claim 56 , wherein the means for selecting phoneme sub-strings further comprises:
means for selecting phoneme strings in the pronunciation data neighborhood for each context within the out-of-vocabulary word;
means for aligning the selected phoneme strings on common phoneme sub-strings; and
means for selecting a phoneme sub-string for each context from the common phoneme sub-strings.
58. The apparatus of claim 57 further comprising:
means for merging the phoneme sub-strings for adjacent contexts when two phoneme sub-strings overlap.
59. An apparatus comprising:
means for storing pronunciation data for an out-of-vocabulary word, the pronunciation data comprising phoneme sub-strings selected from candidate phoneme strings having pronunciation data associated with orthographic anchors that are close to a vector representation the out-of-vocabulary word in an orthographical vector space defined by a dictionary; and
means for transmitting the pronunciation data for the out-of-vocabulary word as an audible signal.
60. The apparatus of claim 59 , wherein the orthographic space comprises the vector representation of the out-of-vocabulary word and the orthographic anchors corresponding to in-vocabulary words.
61. The apparatus of claim 60 , wherein the representation of the out-of-vocabulary word and the orthographic anchors are feature vectors.
62. The apparatus of claim 59 , wherein the dictionary comprises phoneme strings for in-vocabulary words.