Speech recognition system, training arrangement and method of calculating iteration values for free parameters of a maximum-entropy speech model
View Patent ↗The invention relates to a speech recognition system and a method of calculating iteration values for free parameters λ α ortho(n) of a maximum-entropy speech model MESM with the aid of the generalized-iterative scaling training algorithm in a computer-supported speech recognition system in accordance with the formula λ α ortho(n+1) =G(λ α ortho(n) , m α ortho , . . . ), where n is an iteration parameter, G a mathematical function, α an attribute in the MESM and m α ortho a desired orthogonalized boundary value in the MESM for the attribute α. It is an object of the invention to further develop the system and method so that they make a fast computation of the free parameters λ possible without a change of the original training object. According to the invention this object is achieved in that the desired orthogonalized boundary value m α ortho is calculated by a linear combination of the desired boundary value m α with desired boundary values m β from attributes β that have a larger range than the attribute α. m α and m β are then desired boundary values of the original training object.
1. A method of calculating iteration values for free parameters λ α ortho(n) of a maximum-entropy speech model MESM in a speech recognition system with the aid of the generalized iterative scaling training algorithm, the method comprising the step of
iteratively determining:
λ α ortho(n+1) =G (λ α ortho(n) , m α ortho , . . . )
where:
n: is an iteration parameter;
G: is a mathematical function;
α: is an attribute in the MESM; and
m α ortho : is a desired orthogonalized boundary value in the MESM for the attribute α,
characterized in that the desired orthogonalized boundary value m α ortho is calculated by linearly combining the desired boundary value m α with desired boundary values m β of attributes β that have a larger range than the attribute α.
2. A method as claimed in claim 1 , characterized in that the calculation of the desired orthogonalized boundary value m α ortho for the attribute α=β 0 comprises the following steps:
a) Selecting all the attributes β i with i=1 . . . g in the speech model that have a larger range RW than the attribute α=β 0 and include the latter;
b) Calculating desired boundary values mβi for the attributes β i with i=0 . . . g;
c) Sorting the attributes β i with i=0 . . . g according to their RW;
d) Selecting one of the attributes β i having the largest RW;
e) Checking whether there are other attributes β k which include the attribute β i and have a larger RW than the selected attribute β i ;
f1) If so, defining a parameter X as a linear combination of the orthogonalized boundary values m βk ortho calculated in step g) during the last run of the steps e) to g) for all the attributes β i that have a larger range and are determined in the most recently run step e);
f2) If not, defining the parameters X to X=0;
g) Calculating the desired orthogonalized boundary value m βk ortho for the attribute β i by arithmetically combining the desired boundary value mβ i with a parameter X; and
h) Repeating the steps e) to g) for the attribute β i −1 whose RW is smaller than or equal to the RW of the attribute β i until the desired orthogonalized boundary value m β0 ortho =m α ortho with i=0 has been calculated in step g).
3. A method as claimed in claim 2 , characterized in that the calculation of the parameter X in step f1) is made according to the following formula:
X
=
∑
k
m
β
k
ortho
.
4. A method as claimed in claim 3 , characterized in that the calculation of the desired orthogonalized boundary value m βi ortho is made in step g) according to the following formula:
m βi ortho =m βi −X.
5. A method as claimed in claim 2 , characterized in that the calculation of the desired boundary values m βi for the attributes β i with i=0, . . . , g is made in step b) by respectively calculating the frequency N(β i ), with which the attribute β i occurs in a training corpus and by subsequently smoothing the calculated frequency value N(β i ).
6. A method as claimed in claim 5 , characterized in that the calculation of the frequency N(β i ) is made by applying a binary attribute function fβ i to the training corpus where fβ i is defined as:
f
β
i
(
h
,
w
)
f
β
i
(
h
,
w
)
=
{
1
if
β
i
fits
in
the
word
sequence
(
h
,
w
)
otherwise
0
and where ƒ βi (h,w) indicates whether the attribute β i correctly describes a pattern predefined by the word sequence (h,w).
7. A method as claimed in claim 1 , characterized in that the mathematical function G has as a further variable the magnitude of a convergence step t α ortho with:
t α ortho =1 /M ortho
where
M ortho : represents for binary functions ƒ α ortho the maximum number of functions which yield the value 1 for the same argument (h,w).
8. A method as claimed in claim 7 , characterized in that the attribute function ƒ α ortho is calculated by linearly combining an attribute function ƒ α with orthogonalized attribute functions ƒ β ortho is calculated from attributes β that have a larger range than the attribute α.
9. A method as claimed in claim 8 , characterized in that the calculation of the orthogonalized attribute function ƒ α ortho for the attribute α=β 0 comprises the following steps:
a) Selecting all the attributes β i with i=1 . . . g in the speech model that have a larger range RW than the attribute α=β 0 and include the latter;
b) Calculating boundary values fβ i for the attributes β i with i=0 . . . g;
c) Sorting the attributes β i with i=0 . . . g according to their RW;
d) Selecting one of the attributes β i having the largest RW;
e) Checking whether there are other attributes βk which include the attribute β i and have a larger RW than the selected attribute β i ;
f1) If so, defining a function F as a linear combination of the orthogonalized attribute function ƒ βk ortho calculated in step g) during the last run of the steps e) to g) for all the attributes β k that have a larger range determined in the most recently run step e);
f2) If not, defining the function F to F=0;
g) Calculating the orthogonalized attribute function ƒ βk ortho for the attribute βi by arithmetically combining the attribute function fβ i with the function F; and
h) Repeating the steps e) to g) for the attribute β i −1 whose range is smaller than or equal to the range of the attribute β i until the orthogonalized attribute function ƒ β0 ortho =ƒ α ortho with i=0 has been calculated in step g).
10. A method as claimed in claim 9 , characterized in that the calculation of the function F in step f1) is made according to the following formula:
F
=
∑
k
f
β
k
ortho
.
11. A method as claimed in claim 9 , characterized in that the calculation of the orthogonalized attribute function ƒ βi ortho in step g) is made according to the following formula:
ƒ βi ortho =ƒ βi −F.
12. A method as claimed in claim 1 , characterized in that the mathematical function G has the following form:
λ
α
ortho
(
n
+
1
)
=
G
(
λ
α
ortho
(
n
)
,
m
α
ortho
,
…
)
=
λ
α
ortho
(
n
)
+
t
α
ortho
·
log
(
[
t
α
ortho
·
m
α
ortho
+
b
α
]
[
t
α
ortho
·
m
α
ortho
(
n
)
+
b
α
]
·
1
-
∑
γ
[
t
γ
ortho
·
m
γ
ortho
(
n
)
+
b
γ
]
1
-
∑
γ
[
t
γ
ortho
·
m
γ
ortho
+
b
γ
]
)
where:
α: refers to a just considered attribute;
γ: refers to all the attributes in the speech model;
t α ortho , t γ ortho : refer to the size of the convergence step with t α ortho =t γ ortho =1/M ortho with
M
ortho
=
max
(
h
,
w
)
(
∑
β
f
β
ortho
(
h
,
w
)
)
;
where M ortho for binary functions ƒ β ortho represents the maximum number of functions which yield the value 1 for the same argument (h,w);
m α ortho , m γ ortho : refers to desired orthogonalized boundary values in the MESM for the attributes α and γ;
m α ortho(n) , m γ ortho(n) : refers to iterative approximate values for the desired boundary values m α ortho , m γ ortho(n) ; and
bα and bγ: refer to constants.
13. A method as claimed in claim 1 , characterized in that the mathematical function has the following form:
λ
α
ortho
(
n
+
1
)
=
G
(
λ
α
ortho
(
n
)
,
m
α
ortho
,
,
…
)
=
λ
α
ortho
(
n
)
+
t
α
ortho
·
log
(
m
α
ortho
m
α
ortho
(
n
)
·
1
-
∑
β
∈
Ai
(
n
)
(
t
β
·
m
β
ortho
(
n
)
)
1
-
∑
β
∈
Ai
(
n
)
(
t
β
·
m
β
ortho
)
)
where:
n: represents the iteration parameter;
A i(n) : represents an attribute group A i(n) with 1≦i≦m selected in the n th iteration step;
α: represents a just considered attribute from the just selected attribute group A i(n) ;
β: represents all the attributes of the attribute group Ai(n);
t α ortho , t β ortho : represents the size of a convergence step with t α ortho =t β ortho =1/M i(n) ortho with
M
i
(
n
)
ortho
=
max
(
h
,
w
)
(
∑
β
∈
Ai
(
n
)
f
β
ortho
(
h
,
w
)
)
where M i(n) ortho represents for binary functions ƒ β ortho the maximum number of functions from the attribute group A i(n) , which yield the value 1 for the same argument (h,w);
m α ortho , m β ortho : represent desired orthogonalized boundary values in the MESM for the attributes α and β respectively;
m α ortho(n) , m β ortho(n) : represent iterative approximate values for the desired boundary values m α ortho , m β ortho ;
where the selection of the group=A i(n) of attributes α, whose associated parameters λ α ortho are adapted to a current iteration step is made either cyclically or according to a predefined criterion.
14. A speech recognition system comprising: a recognition device for recognizing the semantic content of an acoustic signal captured and rendered available by a microphone, more particularly a speech signal, by mapping parts of this signal onto predefined recognition symbols as they are offered by the implemented maximum-entropy speech model MESM, and for generating output signals which represent the recognized semantic content; and a training system for adapting the MESM to recurrent statistical patterns in the speech of a certain user of the speech recognition system; characterized in that the training system calculates free parameters λ in the MESM in accordance with the method as claimed in claim 1 .
15. A training system for adapting the maximum-entropy speech model MESM in a speech recognition system to recurrent statistical patterns in the speech of a certain user of this speech recognition system, characterized in that the training system calculates free parameters λ in the MESM in accordance with the method as claimed in claim 1 .