Procedure for processing noisy speech signals, and apparatus and computer program therefor
Provided are a noise state determination method and an apparatus and a computer readable recording medium therefor. A noisy speech signal processing method according to the present invention includes calculating a transformed spectrum by transforming an input noisy speech signal to a frequency domain; calculating a smoothed magnitude spectrum by reducing magnitude differences of the transformed spectrum between neighboring frames; calculating a search spectrum which represents an estimated noise component of the smoothed magnitude spectrum; and calculating an identification ratio which represents a ratio of a noise component included in the input noisy speech signal, by using the smoothed magnitude spectrum and the search spectrum. Since a small amount of calculation is required and a large-capacity memory is not required, the present invention may be easily implemented as hardware or software. Also, since an adaptive operation is performed with respect to each frequency sub-band, the accuracy of determining a noise state may be improved.
1. A noise state determination method of a noisy speech signal, the method comprising:
calculating a transformed spectrum by transforming an input noisy speech signal to a frequency domain;
calculating a smoothed magnitude spectrum by reducing magnitude differences of the transformed spectrum between neighboring frames;
calculating a search spectrum which represents an estimated noise component of the smoothed magnitude spectrum; and
calculating an identification ratio which represents a ratio of a noise component included in the input noisy speech signal, by using the smoothed magnitude spectrum and the search spectrum;
wherein a whole frequency range of the frequency domain is divided into subbands, and the calculating of the search spectrum and the calculating of the identification ratio are performed on a sub-band basis, and
wherein the search spectrum is calculated by using Equation E-1
T i,j ( f )=κ( j )· U i−1,j ( f )+(1−κ( j ))· S i,j ( f ) (E-1)
wherein, i is a frame index, j (0≦j<J<L) is a sub-band index obtained by dividing a predetermined frequency range 2 L by a sub-band size (=2 L−J ), J and L are natural numbers which respectively determining a total number of the sub-bands and the predetermined frequency range, T i,j (f) is the search spectrum, S i,j (f) is the smoothed magnitude spectrum, U i−1,j (f) is a weighted spectrum indicating a spectrum having a smaller magnitude between the search spectrum and the smoothed magnitude spectrum of the previous frame, and κ(j) (0<κ(J−1)≦κ(j)≦κ(0)≦1) is a differential forgetting factor.
2. The method of claim 1 , wherein the smoothed magnitude spectrum is calculated by using Equation E-2
S i ( f )=α s S i− ( f )+(1−α s )| Y i ( f )| (E-2)
Here, i is a frame index, f is a frequency, S i−1 (f) and S i (f) respectively are smoothed magnitude spectrums of (i−1)th and ith frames, Y i (f) is a transformed spectrum of the ith frame, and α s is a smoothing factor.
3. The method of claim 2 , wherein the calculating of the search spectrum comprises calculating the search spectrum of a current frame by using only the search spectrum of a previous frame or using the smoothed magnitude spectrum of the current frame and a spectrum having a smaller magnitude between the search spectrum and the smoothed magnitude spectrum of the previous frame.
4. The method of claim 3 , wherein the search spectrum is calculated by using Equation E-3
T
i
,
j
(
f
)
=
{
κ
(
j
)
·
U
1
-
1
,
j
(
f
)
+
(
1
-
κ
(
j
)
)
·
S
i
,
j
(
f
)
,
if
S
i
,
j
(
f
)
>
S
i
-
1
,
j
(
f
)
T
i
-
1
,
j
(
f
)
otherwise
.
(
E
-
3
)
5. The method of claim 3 , wherein the search spectrum is calculated by using Equation E-4
T
i
,
j
(
f
)
=
{
T
i
-
1
,
j
(
f
)
,
if
S
i
,
j
(
f
)
>
S
i
-
1
,
j
(
f
)
κ
(
j
)
·
U
i
-
1
,
j
(
f
)
+
(
1
-
κ
(
j
)
)
·
S
i
,
j
(
f
)
,
otherwise
.
(
E
-
4
)
6. The method of claim 3 , wherein the differential forgetting factor is inversely proportional to the sub-band index.
7. The method of claim 6 , wherein the differential forgetting factor is represented as shown in Equation E-5
κ
(
j
)
=
J
κ
(
0
)
-
j
(
κ
(
0
)
-
κ
(
J
-
1
)
)
J
(
E
-
5
)
Here, 0<κ(J−1)≦κ(j)≦κ(0)≦1.
8. The method of claim 3 , wherein the identification ratio is calculated by using Equation E-6
ϕ
i
(
j
)
=
∑
f
=
j
·
SB
f
=
j
+
1
·
SB
min
(
T
i
,
j
(
f
)
,
S
i
,
j
(
f
)
)
∑
f
=
j
·
SB
f
=
j
+
1
·
SB
S
i
,
j
(
f
)
(
E
-
6
)
Here, SB is a sub-band size and min(a, b) indicates a smaller value of the values a and b.
9. The method of claim 8 , wherein the weighted spectrum is defined as shown in Equation E-7
U i,j ( f )=φ i ( j )·S i,j ( f ) (E-7).
10. The method of claim 8 , wherein, if the identification ratio is larger than a predetermined threshold value, a sub-band is determined as a noise-like sub-band, and, if the identification ratio is equal to or smaller than the threshold value, a sub-band is determined as a speech-like sub-band.
11. The method of claim 1 , wherein the transformed spectrum is calculated by performing Fourier transformation.
12. A method of processing a noisy speech signal of a time domain, the method comprising:
generating a transformed signal by performing Fourier transformation on the noisy speech signal;
performing forward searching so as to calculate a search signal which represents an estimated noise component of the noisy speech signal; and
determining a noise state of the noisy speech signal by using the transformed signal and the search signal,
wherein the search signal is calculated by applying a differential forgetting factor to the transformed signal of a current frame and a signal having a smaller magnitude between the search signal and the transformed signal of a previous frame,
wherein the search signal is calculated by using a smoothed signal which is calculated by reducing magnitude differences of the noisy speech signal between neighboring frames,
wherein the search signal is calculated with respect to a plurality of sub-bands obtained by dividing a whole frequency range of a frequency domain, and wherein the search signal is calculated by using Equation E-8
T i,j ( f )=κ( j )· U i−1,j ( f )+(1−κ( j ))· S i,j ( f ) (E-8)
wherein, i is a frame index, j (0≦j<J<L) is a sub-band index obtained by dividing a predetermined frequency range 2 L by a sub-band size (=2 L−J ), J and L are natural numbers which respectively determine a total number of the sub-bands and the predetermined frequency range, T i,j (f) is the search signal, S i,j (f) is the smoothed signal, U i−1,j (f) is a weighted signal indicating a signal having a smaller magnitude between the search signal and the smoothed signal of the previous frame, and κ(j) (0<κ(J−1)≦κ(j)≦κ(0)≦1) is the differential forgetting factor.
13. The method of claim 12 , wherein the forgetting factor is a differential forgetting factor which has a smaller value in a high-frequency band than in a low-frequency band.
14. The method of claim 12 , wherein the search signal is identical to the search signal of the previous frame in a period where a magnitude of the transformed signal increases.
15. The method of claim 12 , wherein the search signal is identical to the search signal of the previous frame in a period where a magnitude of the transformed signal decreases and is larger than the magnitude of the search signal.
16. A noise state determination apparatus of a noisy speech signal, the apparatus comprising:
a transformation unit for calculating a transformed spectrum by transforming an input noisy speech signal to a frequency domain;
a smoothing unit for calculating a smoothed magnitude spectrum by reducing magnitude differences of the transformed spectrum between neighboring frames;
a forward searching unit for calculating a search spectrum which represents an estimated noise component of the smoothed magnitude spectrum; and
a noise state determination unit for calculating an identification ratio which represents a ratio of a noise component included in the input noisy speech signal, by using the smoothed magnitude spectrum and the search spectrum.
17. An apparatus for processing a noisy speech signal of a time domain, wherein the apparatus comprising:
a transformation unit configured to generates a transformed signal by performing Fourier transformation on the noisy speech signal;
a smoothing unit for calculating a smoothed signal by reducing magnitude differences of the transformed signal between neighboring frames;
a forward searching unit configured to performs forward searching so as to calculate a search signal which represents an estimated noise component of the transformed signal; and
a noise state determination unit configured to determine a noise state of the noisy speech signal by using the transformed signal and the search signal, and
wherein the search signal is calculated by applying a differential forgetting factor to the transformed signal of a current frame and a signal having a smaller magnitude between the search signal and the transformed signal of a previous frame,
wherein a whole frequency range of a frequency domain is divided into subbands, and the calculating of the search signal and the calculating of the noise state are performed on a sub-band basis, and
wherein the search signal is calculated by using Equation E-10
T i,j ( f )=κ( j )· U i−1,j ( f )+(1−κ( j ))· S i,j ( f ) (E-10)
wherein, i is a frame index, j (0≦j<J<L) is a sub-band index obtained by dividing a predetermined frequency range 2 L by a sub-band size (=2 L−j ), J and L are natural numbers which respectively determining a total number of the sub-bands and the predetermined frequency range, T i,j (f) is the search signal, S i,j (f) is the smoothed signal, U i−,j (f) is a weighted signal indicating a signal having a smaller magnitude between the search signal and the smoothed signal of the previous frame, and κ(j) (0<κ(J−1)≦κ(j)≦κ(0)≦1) is the differential forgetting factor.
18. A non-transitory computer readable recording medium having recorded thereon a computer program for executing a noise state determination method of a noisy speech signal, the method comprising:
calculating a transformed spectrum by transforming an input noisy speech signal to a frequency domain;
calculating a smoothed magnitude spectrum by reducing magnitude differences of the transformed spectrum between neighboring frames;
calculating a search spectrum which represents an estimated noise component of the smoothed magnitude spectrum; and
calculating an identification ratio which represents a ratio of a noise component included in the input noisy speech signal, by using the smoothed magnitude spectrum and the search spectrum,
wherein a whole frequency range of the frequency domain is divided into subbands, and the calculating of the search spectrum and the calculating of the identification ratio are performed on a subband basis, and
wherein the search spectrum is calculated by using Equation E-11
T i,j ( f )=κ( j )· U i−1,j ( f )+(1−κ( j ))· S i,j ( f ) (E-11)
wherein, i is a frame index, j (0≦j<J<L) is a sub-band index obtained by dividing a predetermined frequency range 2 L by a subband size (=2 L−J ), J and L are natural numbers which respectively determining a total number of the sub-bands and the predetermined frequency range, T i,j (f) is the search spectrum, S i,j (f) is the smoothed magnitude spectrum, U i−1,j is a weighted spectrum indicating a spectrum having a smaller magnitude between the search spectrum and the smoothed magnitude spectrum of the previous frame, and κ(j) (0<κ(J−1)≦κ(j)≦κ(0)≦1) is a differential forgetting factor.
19. A non-transitory computer readable recording medium having recorded thereon a computer program for executing a method of processing a noisy speech signal of a time domain, the method comprising:
generating a transformed signal by performing Fourier transformation on the noisy speech signal;
generating a smoothed signal by reducing magnitude differences of the transformed signal between neighboring frames;
performing forward searching so as to calculate a search signal which represents an estimated noise component of the transformed signal; and
determining a noise state of the noisy speech signal by using the transformed signal and the search signal,
wherein the search signal is calculated by applying a differential forgetting factor to the transformed signal of a current frame and a signal having a smaller magnitude between the search signal and the transformed signal of a previous frame,
a whole frequency range of a frequency domain is divided into subbands, and the generating of the search signal and the determining of the noise state are performed on a subband basis, and
the search signal is calculated by using Equation E-12
T i,j ( f )=κ( j )· U i−1,j ( f )+(1−κ( j ))· S i,j ( f ) E-12
wherein, i is a frame index, j (0≦j<J<L) is a sub-band index obtained by dividing a predetermined frequency range 2 L by a sub-band size (=2 L−J ), J and L are natural numbers which respectively determining a total number of the sub-bands and the predetermined frequency range, T i,j (f) is the search signal, S i,j (f) is the smoothed signal, U i−1,j (f) is a weighted signal indicating a signal having a smaller magnitude between the search signal and the smoothed signal of the previous frame, and κ(j) (0<κ(J−1)≦κ(j)≦κ(0)≦1) is the differential forgetting factor.