Noise reduction for automatic speech recognition
View Patent ↗Disclosed herein is a noise reduction method for automatic speech recognitionl. A noise reduction method for automatic speech recognition, including: computing a magnitude spectrum of a noisy speech containing a clean speech to be recognized and noise affecting the clean speech; computing a power spectrum of the noisy speech; computing an estimate of a power spectrum of the clean speech; computing an estimate of a power spectrum of the noise; computing an estimate of an a priori signal-to-noise ratio as a function of the estimate of the power spectrum of the clean speech and the estimate of the power spectrum of the noise; computing an estimate of an a posteriori signal-to-noise ratio as a function of the power spectrum of the noisy speech and the estimate of the power spectrum of the noise; computing an attenuation gain as a function of the estimate of the a priori signal-to-noise ratio and the estimate of the a posteriori signal-to-noise ratio; and computing an estimate of a magnitude spectrum of the clean speech as a function of the magnitude spectrum of the noisy speech and the attenuation gain. Computing the estimates of the a priori and the a posteriori signal-to-noise ratios includes computing a noise weighting factor for weighting the estimate of the power spectrum of the noise in the computation of the estimates of the a priori and the a posteriori signal-to-noise ratios; computing a spectral flooring factor for flooring the estimates of the a priori and the a posteriori signal-to-noise ratios; and computing the estimates of the a priori and the a posteriori signal-to-noise ratios also as a function of the noise weighting factor and the spectral flooring factor.
1. A noise reduction method for automatic speech recognition, including:
computing a magnitude spectrum (|Y k (m)|) of a noisy speech containing a clean speech to be recognized and noise affecting the clean speech;
computing a power spectrum (|Y k (m)| 2 ) of the noisy speech;
computing an estimate (|{circumflex over (X)} k (m)| 2 ) of a power spectrum of the clean speech;
computing an estimate (|{circumflex over (D)} k (m)| 2 ) of a power spectrum of the noise;
computing an estimate ({circumflex over (ξ)} k (m)) of an a priori signal-to-noise ratio as a function of the estimate (|{circumflex over (X)} k (m)| 2 ) of the power spectrum of the clean speech and the estimate (|{circumflex over (D)} k (m)| 2 ) of the power spectrum of the noise;
computing an estimate ({circumflex over (γ)} k (m)) of an a posteriori signal-to-noise ratio as a function of the power spectrum (|Y k (m)| 2 ) of the noisy speech and the estimate (|{circumflex over (D)} k (m)| 2 ) of the power spectrum of the noise;
computing an attenuation gain (G k (m)) as a function of the estimate ({circumflex over (ξ)} k (m)) of the a priori signal-to-noise ratio and the estimate ({circumflex over (γ)} k (m)) of the a posteriori signal-to-noise ratio;
computing an estimate (|{circumflex over (X)} k (m)|) of a magnitude spectrum of the clean speech as a function of the magnitude spectrum (|Y k (m)|) of the noisy speech and the attenuation gain (G k (m));
characterized in that computing the estimates ({circumflex over (ξ)} k (m), {circumflex over (γ)} k (m)) of the a priori and the a posteriori signal-to-noise ratios includes:
computing a noise weighting factor (α(m)) for weighting the estimate (|{circumflex over (D)} k (m)| 2 ) of the power spectrum of the noise in the computation of the estimates ({circumflex over (ξ)} k (m), {circumflex over (γ)} k (m)) of the a priori and the a posteriori signal-to-noise ratios;
computing a spectral flooring factor (β(m)) for flooring the estimates ({circumflex over (ξ)} k (m), {circumflex over (γ)} k (m)) of the a priori and the a posteriori signal-to-noise ratios; and
computing the estimates ({circumflex over (ξ)} k (m), {circumflex over (γ)} k (m)) of the a priori and the a posteriori signal-to-noise ratios also as a function of the noise weighting factor (α(m)) and the spectral flooring factor (β(m)).
2. A noise reduction method as claimed in claim 1 , wherein the noise weighting factor (α(m)) and the spectral flooring factor (β(m)) are computed as a function of a global signal-to-noise ratio (SNR(m)).
3. A noise reduction method as claimed in claim 2 , wherein the noise weighting factor (α(m)) versus the global signal-to-noise ratio (SNR(m)) has a first substantially constant value when the global signal-to-noise ratio (SNR(m)) is lower than a first threshold, a second substantially constant value lower than the first substantially constant value when the global signal-to-noise ratio (SNR(m)) is higher than a second threshold, and decreasing values when the global signal-to-noise ratio (SNR(m)) ranges between the first and the second thresholds.
4. A noise reduction method as claimed in claim 3 , wherein the noise weighting factor (α(m)) decreases linearly when the global signal-to-noise ratio (SNR(m)) ranges between the first and the second thresholds.
5. A noise reduction method as claimed in claim 2 , wherein the spectral flooring factor (β(m)) versus the global signal-to-noise ratio (SNR(m)) has a first substantially constant value when the global signal-to-noise ratio (SNR(m)) is lower than a first threshold, a second substantially constant value higher than the first substantially constant value when the global signal-to-noise ratio (SNR(m)) is higher than a second threshold, and increasing values when the global signal-to-noise ratio (SNR(m)) ranges between the first and the second thresholds.
6. A noise reduction method as claimed in claim 5 , wherein the spectral flooring factor (β(m)) increases linearly when the global signal-to-noise ratio (SNR(m)) ranges between the first and the second thresholds.
7. A noise reduction method as claimed in claim 1 , wherein the estimate ({circumflex over (γ)} k (m)) of the a posteriori signal-to-noise ratio is computed as follows:
γ
^
k
(
m
)
=
max
(
Y
k
(
m
)
2
α
(
m
)
D
^
k
(
m
)
2
-
1
,
β
(
m
)
)
+
1
wherein:
{circumflex over (γ)} k (m) is the estimate of the a posteriori signal-to-noise ratio for the k-th spectral line;
|Y k (m)| 2 is the k-th spectral line of the power spectrum of the noisy speech;
|{circumflex over (D)} k (m)| 2 is the k-th spectral line of the estimate of the power spectrum of the noise;
α(m) is the noise weighting factor;
β(m) is the spectral flooring factor;
k indexes spectral lines of the spectra; and
m indexes time windows within which the noisy speech is processed for noise reduction.
8. A noise reduction method as claimed in claim 1 , wherein the estimate ({circumflex over (ξ)} k (m)) of the a priori signal-to-noise ratio is computed as follows:
ξ
^
k
(
m
)
=
max
(
η
X
^
k
(
m
-
1
)
2
α
(
m
)
D
^
k
(
m
-
1
)
2
+
(
1
-
η
)
[
γ
^
k
(
m
)
-
1
]
,
β
(
m
)
)
,
η
(
m
)
∈
[
0
,
1
)
wherein:
{circumflex over (ξ)} k (m) is the estimate of the a priori signal-to-noise ratio for the is k-th spectral line;
{circumflex over (γ)} k (m) is the estimate of the a posteriori signal-to-noise ratio for the k-th spectral line of the power spectrum of the noisy speech;
|{circumflex over (X)} k (m)| 2 is the k-th spectral line of the estimate of the power spectrum of the clean speech;
|{circumflex over (D)} k (m)| 2 is the k-th spectral line of the estimate of the power spectrum of the noise;
α(m) is the noise weighting factor;
β(m) is the spectral flooring factor;
k indexes spectral lines of the spectra; and
m indexes time windows within which the noisy speech is processed for noise reduction.
9. A noise reduction method as claimed in claim 1 , wherein the attenuation gain (G k (m)) is computed as follows:
G
k
(
m
)
=
ξ
^
k
(
m
)
1
+
ξ
^
k
(
m
)
exp
(
1
2
∫
v
k
∞
ⅇ
-
t
t
ⅆ
t
)
wherein:
G k (m) is the Ephraim-Malah gain for the k-th spectral line;
{circumflex over (ξ)} k (m) is the estimate of the a priori signal-to-noise ratio for the k-th spectral line;
{circumflex over (γ)} k (m) is the estimate of the a posteriori signal-to-noise ratio for the k-th spectral line;
v
k
(
m
)
=
ξ
^
k
(
m
)
1
+
ξ
^
k
(
m
)
γ
^
k
(
m
)
k indexes spectral lines of the spectra; and
m indexes time windows within which the noisy speech is processed for noise reduction.
10. A noise reduction method as claimed in claim 1 , wherein the estimate (|{circumflex over (D)} k (m)| 2 of the power spectrum of the noise is computed as follows:
D
^
k
(
m
)
2
=
{
λ
D
^
k
(
m
-
1
)
2
+
(
1
-
λ
)
Y
k
(
m
)
2
if
{
Y
k
(
m
)
2
-
D
^
k
(
m
)
2
≤
μ
σ
(
m
)
}
⋀
{
VAD
=
false
}
D
^
k
(
m
-
1
)
2
otherwise
where:
|{circumflex over (D)} k (m)| 2 is the k-th spectral line of the estimate of the power spectrum of the noise;
|Y k (m)| 2 is the k-th spectral line of the power spectrum of the noisy speech;
λ is a weighting factor which controls the update speed of the recursion,
μ is a multiplication factor which controls the allowed dynamics of the noise, and
σ(m) is the noise standard deviation, estimated as follows:
σ 2 ( m )=λσ 2 ( m− 1)+(1−λ(| Y k ( m )| 2 −|{circumflex over (D)} k ( m )| 2 ) 2
11. A noise reduction method as claimed in claim 2 , wherein the global signal-to-noise ratio (SNR(m)) is computed as follows:
SNR
(
m
)
=
10
log
10
(
∑
k
Y
k
(
m
)
2
∑
k
D
^
k
(
m
)
2
)
where:
SNR(m) is the global signal-to-noise ratio;
|{circumflex over (D)} k (m)| 2 is the k-th spectral line of the estimate of the power spectrum of the noise; and
|Y k (m)| 2 is the k-th spectral line of the power spectrum of the noisy speech.
12. An automatic speech recognition system including a noise reduction system configured to implement the method according to claim 1 .
13. A computer program product comprising a computer program code able, when loaded in a processing system, to implement the method according to claim 1 .