Speech recognition system and method for generating a mask of the system
The speech recognition system of the present invention includes: a sound source separating section which separates mixed speeches from multiple sound sources; a mask generating section which generates a soft mask which can take continuous values between 0 and 1 for each separated speech according to reliability of separation in separating operation of the sound source separating section; and a speech recognizing section which recognizes speeches separated by the sound source separating section using soft masks generated by the mask generating section.
1. A speech recognition system comprising:
multiple sound sources;
a sound source separating section which separates mixed speeches from the multiple sound sources; and
at least one processor configured to:
generate a soft mask which can take continuous values between 0 and 1 for each separated speech according to reliability of separation in separating operation of the sound source separating section, and
recognize speeches separated by the sound source separating section using the soft masks,
wherein the reliability of separation R(f,t) is defined as
R
(
f
,
t
)
=
S
^
m
(
f
,
t
)
+
BN
(
f
,
t
)
Y
m
(
f
,
t
)
,
where Y is a sum of a speech Ŝ m , a background noise BN and a leak.
2. A speech recognition system according to claim 1 , wherein the soft masks are determined using a sigmoid function
1/(1+exp(−a(R−b))
where R represents reliability of separation and a and b represent constants.
3. A speech recognition system according to claim 1 , wherein the soft masks are determined using a probability density function of a normal distribution, which has a variable R which represents reliability of separation.
4. A method for generating a soft mask for a speech recognition system, the method comprising:
separating, at a sound source separating section of the speech recognition system, mixed speeches from multiple sound sources;
generating, at a mask generating section of the speech recognition system, a soft mask which can take continuous values between 0 and 1 for each separated speech according to reliability of separation in separating operation of the sound source separating section;
recognizing, at a speech recognizing section of the speech recognition system, speeches separated by the sound source separating section using soft masks generated by the mask generating section, the soft mask being determined using a function of the reliability of separation, which has at least one parameter;
determining a search space of said at least one parameter;
obtaining a speech recognition rate of the speech recognition system while changing a value of the speech recognition system in the search space; and
setting the value which maximizes a speech recognition rate of the speech recognition system to said at least one parameter,
wherein the reliability of separation R(f,t) is defined as
R
(
f
,
t
)
=
S
^
m
(
f
,
t
)
+
BN
(
f
,
t
)
Y
m
(
f
,
t
)
,
where Y is a sum of a speech Ŝ m , a background noise BN and a leak.
5. A method for generating a soft mask for a speech recognition system, the method comprising:
separating, at a sound source separating section of the speech recognition system, mixed speeches from multiple sound sources;
generating, at a mask generating section of the speech recognition system, a soft mask which can take continuous values between 0 and 1 for each separated speech according to reliability of separation in separating operation of the sound source separating section;
recognizing, at a speech recognizing section of the speech recognition system, speeches separated by the sound source separating section using soft masks generated by the mask generating section, the soft mask being determined using a function of the reliability of separation, which has at least one parameter;
obtaining a histogram of the reliability of separation; and
determining a value of said at least one parameter from a form of the histogram of the reliability of separation,
wherein the reliability of separation R(f,t) is defined as
R
(
f
,
t
)
=
S
^
m
(
f
,
t
)
+
BN
(
f
,
t
)
Y
m
(
f
,
t
)
,
where Y is a sum of a speech Ŝ m , a background noise BN and a leak.
6. A method for generating a soft mask for a speech recognition system according to claim 5 , wherein assuming that
μ1 and μ2 (μ1<μ2)
indicate mean values and
σ1 and σ2
indicate standard deviations and R indicates reliability of separation, the mean values and standard deviations
μ1, μ2, σ1 and σ2
are estimated by fitting the histogram of reliability of separation R with a first probability density function of normal distribution f 1 (R) which has
(μ1,σ1)
and a second probability density function of normal distribution f 2 (R) which has
(μ2,σ2)
and the soft mask is generated using f 1 (R), f 2 (R),
μ1 and μ2.
7. A method for generating a soft mask for a speech recognition system according to claim 6 , wherein assuming that a value of the soft mask is S(R) and f(R)=f 1 (R)+f 2 (R),
S(R)=0 when R<μ1,
S(R)=f 2 (R)/f(R) when μ1≦R≦μ2
S(R)=1 when μ2<R.
8. A method for generating a soft mask for a speech recognition system according to claim 6 , wherein assuming that a value of the soft mask is S(R),
f
1
′
(
R
)
=
1
2
π
σ
2
when
R
<
μ
1
,
f
1
′
(
R
)
=
f
1
(
R
)
when
μ
1
≤
R
,
f
2
′
(
R
)
=
f
2
(
R
)
when
R
<
μ
2
,
f
2
′
(
R
)
=
1
2
π
σ
2
when
μ
2
≤
R
,
and
f
′
(
R
)
=
f
1
′
(
R
)
+
f
2
′
(
R
)
,
S
M
(
R
)
=
f
2
′
(
R
)
f
′
(
R
)
,
wherein SM(R) represents a soft missing feature mask (MFM).
9. A method for generating a soft mask for a speech recognition system according to claim 6 , wherein a value of R at the intersection of f 1 (R) and f 2 (R) which satisfies
μ1<R<μ2
is set to b and a is determined such that
1/(1+exp(− a ( R−b ))
is fit to
f2(R)/f(R)
and the value of the missing feature mask (MFM) S(R) is determined by
S ( R )=1/(1+exp(− a ( R−b )).
10. A speech recognition system according to claim 1 , wherein assuming that
μ1 and μ2 (μ1<μ2)
indicate mean values,
σ1 and σ2
indicate standard deviations, and R indicates the reliability of separation, the mean values and standard deviations
μ1, μ2, σ1 and σ2 are estimated by fitting a histogram of the reliability of separation R with a first probability density function of normal distribution f 1 (R) which has
(μ1,σ1)
and a second probability density function of normal distribution f 2 (R) which has
(μ2,σ2)
and the soft mask is generated using f 1 (R), f 2 (R),
μ1 and μ2.