Discriminative training for speaker and speech verification
View Patent ↗A method for discriminatively training acoustic models is provided for automated speaker verification (SV) and speech (or utterance) verification (UV) systems. The method includes: defining a likelihood ratio for a given speech segment, whose speaker identity (for SV system) or linguist identity (for UV system) is known, using a corresponding acoustic model, and an alternative acoustic model which represents all other speakers (in SV) or all other linguist identities (in UV); determining an average likelihood ratio score for the likelihood ratio scores over a set of training utterances (referred to as true data set) whose speaker identities (for SV) or linguist identities (for UV) are the same; determining an average likelihood ratio score for the likelihood ratio scores over a competing set of training utterances which excludes the speech data in the true data set (referred to as competing data set); and optimizing a difference between the average likelihood ratio score over the true data set and the average likelihood ratio score over the competing data set, thereby improving the acoustic model.
1. A method for discriminatively training acoustic models for automated speech verification, comprising:
defining a likelihood ratio for a given speech segment X having a known linguist identity W, using an acoustic model which represents W and an alternative acoustic model which represents linguist identities other than W;
determining an average likelihood ratio score for the likelihood ratio scores over a set of training utterances whose linguist identities are the same, W;
determining an average likelihood ratio score for the likelihood ratio scores over a competing set of training utterances whose linguist identities are not W; and
optimizing a difference between the average likelihood ratio score over the set of training utterances and the average likelihood ratio score over the competing set of training utterances, thereby improving the acoustic model.
2. The method of claim 1 wherein the acoustic models are further defined as hidden Markov models.
3. The method of claim 1 wherein defining a likelihood ratio further comprises using a log-likelihood ratio test, LTR(X), defined as
LRT
(
X
)
=
log
P
(
X
|
λ
W
)
P
(
X
|
λ
W
_
)
=
log
P
(
X
|
λ
W
)
-
log
P
(
X
|
λ
W
_
)
where λ W and λ W represent the acoustic model and the alternative acoustic model, respectively.
4. The method of claim 3 wherein optimizing a difference further comprises defining a training criterion as
Λ
opt
=
arg
max
(
λ
W
,
λ
W
_
)
Q
(
λ
W
,
λ
W
_
)
where Q(λ W ,λ W ) is defined as
Q
(
λ
W
,
λ
W
_
)
=
1
S
T
∑
X
∈
S
T
LRT
(
X
)
-
1
S
C
∑
X
∈
S
C
LRT
(
X
)
-
θ
where S T is the set of training utterances and S C is the competing set of training utterances.
5. The method of claim 3 wherein optimizing a difference further comprises defining a training criterion as
Λ
opt
=
arg
max
(
λ
W
,
λ
W
_
)
Q
1
(
λ
W
,
λ
W
_
)
where Q 1 (λ W ,λ W ) is defined as
Q
1
(
λ
W
,
λ
W
_
)
=
-
1
S
T
∑
X
∈
S
T
LRT
(
X
)
+
1
S
C
∑
X
∈
S
C
LRT
(
X
)
-
θ
where S T is the set of training utterances and S C is the competing set of training utterances.
6. The method of claim 5 further comprises solving the training criterion using a generalized probabilistic descent algorithm or a Quickprop optimization algorithm.
7. The method of claim 3 wherein optimizing a difference further comprises defining a training criterion as
Λ
opt
=
arg
max
(
λ
W
,
λ
W
_
)
Q
2
(
λ
W
,
λ
W
_
)
where Q 2 (λ W ,λ W ) is defined as
Q
2
(
λ
W
,
λ
W
_
)
=
1
S
T
+
S
C
(
∑
X
∈
S
T
LRT
(
X
)
-
∑
X
∈
S
C
LRT
(
X
)
)
-
θ
where S T is the set of training utterances and S C is the competing set of training utterances.
8. The method of claim 1 wherein optimizing a difference further comprises maximizing the difference by iteratively adjusting parameters of the acoustic models.
9. A method for discriminatively training acoustic models for automated speaker verification, comprising:
defining a likelihood ratio for a given speech segment X having a known speaker identity, K, using an acoustic model which represents K and an alternative acoustic model which represents speakers other than K;
determining an average likelihood ratio score for the likelihood ratio scores over a set of training utterances which are spoken by speaker K;
determining an average likelihood ratio score for the likelihood ratio scores over a competing set of training utterances which are spoken by speakers other than speaker K; and
optimizing a difference between the average likelihood ratio score over the set of training utterances and the average likelihood ratio score over the competing set of training utterances, thereby improving the acoustic model.
10. The method of claim 9 wherein defining a likelihood ratio further comprises using a log-likelihood ratio test, LRT(X) defined as
LRT
(
X
)
=
log
P
(
X
|
λ
K
)
P
(
X
|
λ
K
_
)
=
log
P
(
X
|
λ
K
)
-
log
P
(
X
|
λ
K
_
)
where λ K and λ K represent the acoustic model and the alternative acoustic model, respectively.
11. The method of claim 10 wherein optimizing a difference further comprises defining a training criterion as
Λ
opt
=
arg
max
(
λ
K
,
λ
K
_
)
Q
(
λ
K
,
λ
K
_
)
where Q(λ K , λ K ) is defined as
Q
(
λ
K
,
λ
K
_
)
=
1
S
T
∑
X
∈
S
T
LRT
(
X
)
-
1
S
C
∑
X
∈
S
C
LRT
(
X
)
-
θ
where S T is the set of training utterances and S C is the competing set of training utterances.
12. The method of claim 10 wherein optimizing a difference further comprises defining a training criterion as
Λ
opt
=
arg
min
(
λ
K
,
λ
K
_
)
Q
(
λ
K
,
λ
K
_
)
where Q(λ K ,λ K ) is defined as
Q
(
λ
K
,
λ
K
_
)
=
-
1
S
T
∑
X
∈
S
T
LRT
(
X
)
+
1
S
C
∑
X
∈
S
C
LRT
(
X
)
-
θ
where S T is the set of training utterances and S C is the competing set of training utterances.
13. The method of claim 12 further comprises solving the training criterion using a generalized probabilistic descent algorithm or a Quickprop optimization algorithm.
14. The method of claim 10 wherein optimizing a difference further comprises defining a training criterion as
Λ
opt
=
arg
min
(
λ
K
,
λ
K
_
)
Q
2
(
λ
K
,
λ
K
_
)
where Q 2 (λ K ,λ K ) is defined as
Q
2
(
λ
K
,
λ
K
_
)
=
1
S
T
+
S
C
(
∑
X
∈
S
T
LRT
(
X
)
-
∑
X
∈
S
C
LRT
(
X
)
)
-
θ
where S T is the set of training utterances and S C is the competing set of training utterances.
15. The method of claim 9 wherein optimizing a difference further comprises maximizing the difference by iteratively adjusting parameters of the acoustic model.