Post tone suppression for speech enhancement
View Patent ↗A method for reducing disturbing tone signals from acoustic echoes or background noises in speech/audio signal enhancement processing comprises decomposing a full-band input microphone speech signal into plurality of sub-band component channel signals by using a filter-bank; disturbing tone signal in each sub-band component channel signal is detected by using parameters such as a second reflection coefficient from LPC analysis, a time domain sharpness parameter, a normalized pitch correlation, and/or a SNR parameter; the energy of the current sub-band component channel signal is reduced by multiplying a reduction gain if the disturbing tone signal is detected in the current sub-band component channel signal; all the component channel signals are summed back to output a tone-reduced full band speech signal.
1. A method for reducing disturbing tone signals from acoustic echoes or background noises in speech/audio signal enhancement processing, the method comprising:
decomposing a full-band input microphone speech signal into plurality of sub-band component channel signals by using a filter-bank;
detecting disturbing tone signal in each sub-band component channel signal wherein the parameters comprise at least one of a second reflection coefficient from LPC analysis, a time domain sharpness parameter, a normalized pitch correlation, and/or a SNR parameter;
reducing the energy of the current sub-band component channel signal by multiplying a reduction gain if the disturbing tone signal is detected in the current sub-band component channel signal;
summing back all the component channel signals to output a tone-reduced full band speech signal.
2. The method of claim 1 , wherein reducing disturbing tone signals happens after acoustic echo cancellation and/or microphone array beamforming processing.
3. The method of claim 1 , wherein the second reflection coefficient obtained from LPC analysis is calculated on the component channel signal with a Levinson-Durbin approach.
4. The method of claim 1 , wherein the time domain sharpness parameter is defined as an average value divided by a maximum value wherein the average value and the maximum value are calculated on a frame of the component channel signal in time domain:
Sharp_ch
=
∑
n
=
0
L
-
1
S
ch
(
n
)
L
·
Max
{
S
ch
(
n
)
,
n
=
0
,
1
,
…
,
L
-
1
}
wherein S ch (n) is the component channel signal, and L is the fame length.
5. The method of claim 1 , wherein the normalized pitch correlation is calculated on the component channel signal in time domain:
Rp_ch
=
∑
n
S
ch
(
n
)
·
S
ch
(
n
-
Pitch
)
[
∑
n
S
ch
(
n
)
2
]
[
∑
n
S
ch
(
n
-
Pitch
)
2
]
wherein Pitch is estimated based on the full band signal (not the component channel signal).
6. The method of claim 1 , wherein the SNR is calculated in dB on a frame of the full band speech signal and/or the sub-band component channel signal.
7. The method of claim 1 , wherein the reduction gain is larger than or equal to zero and smaller than or equal to one.
8. A speech signal processing apparatus comprising:
a processor; and
a computer readable storage medium storing programming for execution by the processor, the programming including instructions to:
decompose a full-band input microphone speech signal into plurality of sub-band component channel signals by using a filter-bank;
detect disturbing tone signal in each sub-band component channel signal wherein the parameters comprise at least one of a second reflection coefficient from LPC analysis, a time domain sharpness parameter, a normalized pitch correlation, and/or a SNR parameter;
reduce the energy of the current sub-band component channel signal by multiplying a reduction gain if the disturbing tone signal is detected in the current sub-band component channel signal;
sum back all the component channel signals to output a tone-reduced full band speech signal.
9. The speech signal processing apparatus of claim 8 , wherein reducing disturbing tone signals happens after acoustic echo cancellation and/or microphone array beamforming processing.
10. The speech signal processing apparatus of claim 8 , wherein the second reflection coefficient obtained from LPC analysis is calculated on the component channel signal with a Levinson-Durbin approach.
11. The speech signal processing apparatus of claim 8 , wherein the time domain sharpness parameter is defined as an average value divided by a maximum value wherein the average value and the maximum value are calculated on a frame of the component channel signal in time domain:
Sharp_ch
=
∑
n
=
0
L
-
1
S
ch
(
n
)
L
·
Max
{
S
ch
(
n
)
,
n
=
0
,
1
,
…
,
L
-
1
}
wherein S ch (n) is the component channel signal, and L is the fame length.
12. The speech signal processing apparatus of claim 8 , wherein the normalized pitch correlation is calculated on the component channel signal in time domain:
Rp_ch
=
∑
n
S
ch
(
n
)
·
S
ch
(
n
-
Pitch
)
[
∑
n
S
ch
(
n
)
2
]
[
∑
n
S
ch
(
n
-
Pitch
)
2
]
wherein Pitch is estimated based on the full band signal (not the component channel signal).
13. The speech signal processing apparatus of claim 8 , wherein the SNR is calculated in dB on a frame of the full band speech signal and/or the sub-band component channel signal.
14. The speech signal processing apparatus of claim 8 , wherein the reduction gain is larger than or equal to zero and smaller than or equal to one.