Incrementally regulated discriminative margins in MCE training for speech recognition
View Patent ↗A method and apparatus for training an acoustic model are disclosed. A training corpus is accessed and converted into an initial acoustic model. Scores are calculated for a correct class and competitive classes, respectively, for each token given the acoustic model. From this score a misclassification measure is calculated and then a loss function is calculated from the misclassification measure. The loss function also includes a margin value that varies over each iteration in the training. Based on the calculated loss function the acoustic model is updated, where the loss function with the margin value is minimized. This process repeats until such time as an empirical convergence is met.
1. A method of training an acoustic model in a speech recognition system, comprising:
accessing a training corpus;
using the training corpus to calculate model parameter values for an initial form of an acoustic model;
a processor iteratively updating model parameter values for the acoustic model, each iteration comprising:
calculating a plurality of scores for each token with regard to a correct class and a plurality of competing classes from a set of model parameter values determined for the acoustic model before the iteration;
determining a value for a loss function based on the calculated scores and a margin wherein the loss function is computed as:
l
r
(
d
r
(
X
r
,
Λ
)
)
=
1
1
+
ⅇ
-
ad
(
X
r
,
Λ
)
+
β
(
I
)
;
where l r (d r (X r ,Λ)) is the value for the loss function, d r (X r , Λ) is a misclassification measure for a token X r determined from the plurality of scores for the token X r , Λ represents the set of model parameter values for the acoustic model, α is a constant, β(I) is the margin, and I is an iteration argument, wherein for at least two different iterations that form part of iteratively updating model parameter values for a same acoustic model, the value of β(I) is different;
updating the set of model parameter values to create a revised set of model parameter values based upon the value of the loss function; and
outputting the a final revised set of model parameter values for the acoustic model.
2. The method of claim 1 further comprising:
iteratively updating the set of model parameter values until an empirical convergence has been met for the set of model parameter values.
3. The method of claim 2 wherein the empirical convergence occurs after four iterations.
4. The method of claim 1 where four iterations are performed for each value of β(I).
5. A system for training an acoustic model comprising:
a training corpus;
a training component; and
wherein the training component is configured to iteratively generate the acoustic model based on the training corpus and a loss function wherein the loss function is defined as
l
r
(
d
r
(
X
r
,
Λ
)
)
=
1
1
+
ⅇ
-
ad
r
(
X
r
,
Λ
)
+
β
(
I
)
;
where d r (X r ,Λ) is a misclassification measure, α is a constant, β(I) is a margin, I is an iteration argument, X r is a string vector, r is an index and Λ represents model parameters of the acoustic model, wherein at least two different iterations represented by two different respective values of the iteration argument use two different values for β(I) in the loss function.
6. The system of claim 5 further comprising:
speaker independent data; and
wherein the training component is configured to use the speaker independent data in conjunction with the training corpus to generate the acoustic model.
7. The system of claim 6 wherein the training component generates a custom acoustic model for each speaker.
8. The system of claim 5 wherein the training component comprises two separate acoustic model training components.
9. The system of claim 5 where four iterations are performed for each value of β(I).
10. The system of claim 5 wherein the training component is configured to iteratively generate the acoustic model until an empirical convergence is achieved.
11. A speech recognition system configured to recognize a spoken utterance, comprising:
a feature extraction module;
a training module;
a search engine;
a lexicon;
a language model;
an acoustic model; and
wherein the acoustic model is iteratively trained based on a training data set and a computed misclassification measure processed through a loss function that includes a margin, wherein the loss function is defined as
l
r
(
d
r
(
X
r
,
Λ
)
)
=
1
1
+
e
-
α
d
r
(
X
r
,
Λ
)
+
β
(
I
)
where d r (X r ,Λ) is a misclassification measure, α is a constant, β(I) is a margin, I is an iteration argument, X r is a string vector, r is an index and Λ represents model parameters of the acoustic model, wherein different iterations use two different values for β(I) in the loss function.
12. The speech recognition system of claim 11 where four iterations are performed for each value of β(I).