Method and apparatus for feature domain joint channel and additive noise compensation
View Patent ↗A method for performing noise adaptation of a target speech signal input to a speech recognition system, where the target speech signal contains both additive and convolutional noises. The method includes estimating an additive noise bias and a convolutional noise bias; in the target speech signal; and jointly compensating the target speech signal for the additive and convolutional noise biases in a feature domain.
1. A method for performing noise adaptation of a target speech signal input to a speech recognition system, said target speech signal containing both additive and convolutional noises, said method comprising:
estimating an additive noise bias and a convolutional noise bias in said target speech signal; and
jointly compensating the target speech signal for the additive and convolutional noise biases in a feature domain,
wherein said target speech signal comprises a plurality of frames, and further wherein said estimating a convolutional noise bias comprises:
determining a speech signal partially adapted to the test environment in additive noise; and
utilizing the partially adapted speech signal to estimate the convolutional noise bias, and
wherein said determining a speech signal partially adapted to the test environment in additive noise comprises determining a first order approximation as:
C
(
H
tar
·
S
+
N
ref
)
=
C
(
H
tar
·
S
+
N
tar
)
+
∂
C
(
H
tar
·
S
+
N
)
∂
N
N
=
N
tar
(
N
ref
-
N
tar
)
,
=
C
(
H
tar
·
S
+
N
tar
)
+
J
N
Δ
N
,
wherein:
C(·) is the cepstral operator;
H tar is a channel frequency response for a test environment;
N ref is an additive noise for a training environment;
N tar is an additive noise for the test environment;
S represents a clean version of the target speech signal;
J
N
is
∂
C
(
H
tar
·
S
+
N
)
∂
N
N
=
N
tar
,
a
noise
compensation
matrix
;
ΔN is (N ref −N tar ); and
C(H tar ·S+N ref ) is said partially adapted speech signal.
2. A method in accordance with claim 1 further comprising determining noise compensation matrix J N as:
J
N
=
F
†
1
H
tar
·
S
+
N
tar
,
where F † is a decorrelating transform and H tar ·S+N tar is the noisy target speech signal.
3. A method in accordance with claim 1 wherein said partially adapted speech signal utilized to estimate the convolutional noise bias of one of said frames of said target speech signal is a partially adapted speech signal determined from a previous said signal frame.
4. A method in accordance with claim 3 wherein said estimating a convolutional noise bias further comprises determining a model H ref ·S+N ref corresponding to said target speech signal in the spectral domain, wherein H ref is a channel frequency response for a training environment.
5. A method in accordance with claim 4 wherein said determining a model corresponding to said target speech signal in the spectral domain comprises utilizing Viterbi decoding to evaluate a most likely partial path.
6. A method in accordance with claim 5 wherein said convolutional noise bias ΔH=H ref −H tar is determined as:
ΔH=J p −1 [C ( H ref ·S+N ref )−C( H tar ·S+N ref )],
wherein:
J p −1 is a matrix inverse, in which
J
P
=
∂
C
(
H
·
S
+
N
ref
)
∂
H
H
=
H
ref
=
F
†
S
H
ref
·
S
+
N
ref
;
and F † is a decorrelating transform.
7. A method in accordance with claim 6 wherein F † is a discrete cosine transform.
8. A method for performing noise adaptation of a target speech signal input to a speech recognition system, said target speech signal containing both additive and convolutional noises, said method comprising:
estimating an additive noise bias and a convolutional noise bias in said target speech signal; and
jointly compensating the target speech signal for the additive and convolutional noise biases in a feature domain,
wherein said target speech signal comprises a plurality of frames, and further wherein said estimating a convolutional noise bias comprises:
determining a speech signal partially adapted to the test environment in additive noise; and
utilizing the partially adapted speech signal to estimate the convolutional noise bias; and
wherein said partially adapted speech signal utilized to estimate the convolutional noise bias of one of said frames of said target speech signal is a partially adapted speech signal determined from a previous said signal frame, said estimating a convolutional noise bias further comprises determining a model corresponding to said target speech signal in the spectral domain,
the method further comprising determining said model in the spectral domain from a model in the cepstral domain utilizing an inverse cepstral operator C −1 (·), wherein C −1 (x)=exp(F −1 x), F is a matrix, F † is a decorrelating transform and the inverse conjugate of F, and FF † =I, where I is an identity matrix.
9. A speech recognition apparatus including noise adaptation of a target speech signal input to said apparatus, said target speech signal containing both additive and convolutional noises, said apparatus configured to:
estimate an additive noise bias and a convolutional noise bias in said target speech signal; and
jointly compensate the target speech signal for the additive and convolutional noise biases in a feature domain,
wherein said target speech signal comprises a plurality of frames, and further wherein to estimate a convolutional noise bias, said apparatus is configured to:
determine a speech signal partially adapted to the test environment in additive noise; and
utilize the partially adapted speech signal to estimate the convolutional noise bias,
wherein to determine a speech signal partially adapted to the test environment in additive noise said apparatus is configured to determine a first order approximation as:
C
(
H
tar
·
S
+
N
ref
)
=
C
(
H
tar
·
S
+
N
tar
)
+
∂
C
(
H
tar
·
S
+
N
)
∂
N
N
=
N
tar
(
N
ref
-
N
tar
)
,
=
C
(
H
tar
·
S
+
N
tar
)
+
J
N
Δ
N
,
Wherein:
C(·) is the cepstral operator;
H tar is a channel frequency response for a test environment;
N ref is an additive noise for a training environment;
N tar is an additive noise for the test environment;
S represents a clean version of the target speech signal;
J
N
is
∂
C
(
H
tar
·
S
+
N
)
∂
N
N
=
N
tar
,
a
noise
compensation
matrix
;
ΔN is (N ref −N tar ); and
C(H tar ·S+N ref ) is said partially adapted speech signal.
10. An apparatus in accordance with claim 9 further configured to determine noise compensation matrix J N as:
J
N
=
F
†
1
H
tar
·
S
+
N
tar
,
where F † is a decorrelating transform and H tar ·S+N tar is the noisy target speech signal.
11. An apparatus in accordance with claim 9 wherein said partially adapted speech signal utilized to estimate the convolutional noise bias of one of said frames of said target speech signal is a partially adapted speech signal determined from a previous said signal frame.
12. An apparatus in accordance with claim 11 wherein to estimate a convolutional noise bias, said apparatus is further configured to determine a model H ref ·S+N ref corresponding to said target speech signal in the spectral domain, wherein H ref is a channel frequency response for a training environment.
13. An apparatus in accordance with claim 12 wherein to determine a model corresponding to said target speech signal in the spectral domain, said apparatus is configured to utilize Viterbi decoding to evaluate a most likely partial path.
14. An apparatus in accordance with claim 13 configured to determine said convolutional noise bias ΔH=H ref −H tar as:
ΔH=J p −1 [C ( H ref ·S+N ref ) −C ( H tar ·S+N ref )] (15)
wherein:
J −1 p is a matrix inverse, in which
J
P
=
∂
C
(
H
·
S
+
N
ref
)
∂
H
H
=
H
ref
=
F
†
S
H
ref
·
S
+
N
ref
;
and F † is a decorrelating transform.
15. An apparatus in accordance with claim 14 wherein F † is a discrete cosine transform.
16. An apparatus in accordance with claim 15 wherein said partially adapted speech signal utilized to estimate the convolutional noise bias of one of said frames of said target speech signal is a partially adapted speech signal determined from a previous said signal frame.
17. An apparatus in accordance with claim 16 wherein to estimate a convolutional noise bias, said apparatus is further configured to determine a model corresponding to said target speech signal in the spectral domain.
18. An apparatus in accordance with claim 17 wherein to determine a model corresponding to said target speech signal in the spectral domain, said apparatus is configured to utilize Viterbi decoding to evaluate a most likely partial path.
19. An apparatus in accordance with claim 18 further configured to determine said model in the spectral domain from a model in the cepstral domain utilizing an inverse cepstral operator C −1 (·), wherein C −1 (x)=exp(F −1 x), F is a matrix, F † is a decorrelating transform and the inverse conjugate of F, and FF † =I, where I is an identity matrix.