Methods and systems for blind dereverberation
View Patent ↗Various embodiments of the present invention are directed to methods for dereverberation of audio generated in a room. In one aspect, a method for dereverberating reverberant digital signals comprises transforming a reverberant digital signal from the time domain into Fourier domain signals using a computing device, each Fourier domain signal corresponding to a subband. For each subband of the Fourier domain signal, the method computes autoregressive model coefficients of the reverberation with the current and previous magnitudes of the Fourier digital signal, and inverse filters the magnitude of the Fourier domain signal using the computing device, based on the autoregressive model coefficients and previous magnitudes of the Fourier digital signal. The method includes inverse transforming the Fourier domain signals with filtered magnitudes into an approximate dereverberated digital signal.
1. A method for dereverberating digital source signals using a computing device, the method comprising:
transforming a reverberant digital signal into Fourier domain signals using the computing device, each Fourier domain signal corresponding to a subband;
for each subband,
computing first or higher order autoregressive model coefficients of the reverberation with current and previous magnitudes of the Fourier domain signal, and
inverse filtering a magnitude of the Fourier domain signal based on the autoregressive model coefficients and previous magnitudes of the Fourier domain signal;
inverse transforming the Fourier domain signals with filtered magnitudes to obtain an approximate dereverberated digital signal;
wherein inverse filtering the magnitude of the Fourier domain signal further comprising computing:
Ŵ k [r]=Ŷ k [r]−â k [r]Ŷ k [r−M]
where
Ŵ k [r] is the filtered magnitude of the Fourier domain signal,
Ŷ k [r] is the magnitude of the Fourier domain signal,
Ŷ k [r−M] is the magnitude of a previous Fourier domain signal, and
â k [r] is the approximate model coefficient.
2. The method of claim 1 further comprising, for each subband converting the Fourier domain signal into polar form.
3. The method of claim 1 wherein transforming the reverberant digital signal from the time domain into Fourier domain signals further comprises applying a short-time Fourier transform on the reverberant digital signal.
4. The method of claim 1 wherein transforming the reverberant digital signal from the time domain into Fourier domain signals further comprises critically and over sampling the subbands.
5. The method of claim 1 wherein computing the autoregressive model coefficients further comprises computing an autocovariance approximation of the magnitudes as follows:
a
^
k
[
r
]
=
a
^
k
(
num
)
[
r
]
a
^
k
(
den
)
[
r
]
where
â k (num) [r]=αâ k (num) [r− 1]+(1−α)( Ŷ k [r ]−{circumflex over (μ)} Ŷ k [r ])( Ŷ k [r−M ]−{circumflex over (μ)} Ŷ k [r ]),
â k (den) [r]=αâ k (den) [r− 1]+(1−α)( Ŷ k [r]−{circumflex over (μ)} Ŷ k [r ]) 2 ,
Ŷ k [r] is the magnitude of the Fourier domain signal,
Ŷ k [r−M] is the magnitude of a previous Fourier domain signal
{circumflex over (μ)} Ŷ k [r]=α{circumflex over (μ)} Ŷ k [r−1]+(1−α)Ŷ k [r], and
α is a parameter that determines the rate at which the estimates of the a k are allowed to adapt.
6. A non-transitory computer readable medium having instructions thereon which, when executed by a computing device, dereverberate a reverberant digital signal into an approximate dereverberated digital signal, the instructions comprising instructions to:
transform a reverberant digital signal into short-time Fourier domain signals, each Fourier domain signal corresponding to a subband;
for each subband:
compute at least first order autoregressive model coefficients of the reverberation with current and previous magnitudes of the Fourier domain signal; and
inverse filter a magnitude of the Fourier domain signal based on the autoregressive model coefficients and previous magnitudes of the Fourier domain signal; and
inverse transform the Fourier domain signals with filtered magnitudes to obtain an approximate dereverberated digital signal;
wherein to inverse filter the magnitude of the Fourier domain signal, the instructions cause the computer device to compute:
Ŵ k [r]=Ŷ k [r]−â k [r]Ŷ k [r−M]
where
Ŵ k [r] is the filtered magnitude of the Fourier domain signal,
Ŷ k [r] is the magnitude of the Fourier domain signal,
Ŷ k [r−M] is the magnitude of a previous Fourier domain signal, and
â k [r] is the approximate model coefficient.
7. The non-transitory computer readable medium of claim 6 , wherein the instructions to cause the computer device to compute the at least first order autoregressive model coefficients of the reverberation further comprise instructions to cause the computer device to compute an autocovariance approximation of the current and previous magnitudes as follows:
a
^
k
[
r
]
=
a
^
k
(
num
)
[
r
]
a
^
k
(
den
)
[
r
]
where
â k (num) [r]=αâ k (num) [r− 1]+(1−α)( Ŷ k [r]−{circumflex over (μ)} Ŷ k [r ])( Ŷ k [r−M]−{circumflex over (μ)} Ŷ k [r ]),
â k (den) [r]=αâ k (den) [r− 1]+(1−α)( Ŷ k [r ]−{circumflex over (μ)} Ŷ k [r ]) 2 ,
Ŷ k [r] is the magnitude of the Fourier domain signal,
Ŷ k [r−M] is the magnitude of a previous Fourier domain signal
{circumflex over (μ)} Ŷ k [r]=α{circumflex over (μ)} Ŷ k [r−1]+(1−α)Ŷ k [r], and
α is a parameter that determines the rate at which the estimates of the a k are allowed to adapt.
8. A system to dereverberate a reverberant digital signal into an approximate dereverberated digital signal, comprising:
one or more computer processors; and
a memory, encoded thereon computer instructions directing the one or more computer processors to:
transform a reverberant digital signal into short-time Fourier domain signals, each Fourier domain signal corresponding to a subband;
for each subband:
compute at least first order autoregressive model coefficients of the reverberation with current and previous magnitudes of the Fourier domain signal; and
inverse filter a magnitude of the Fourier domain signal based on the autoregressive model coefficients and previous magnitudes of the Fourier domain signal; and
inverse transform the Fourier domain signals with filtered magnitudes to obtain an approximate dereverberated digital signal;
wherein to inverse filter the magnitude of the Fourier domain signal, the instructions cause the computer device to compute:
Ŵ k [r]=Ŷ k [r]−â k [r]Ŷ k [r−M]
where
Ŵ k [r] is the filtered magnitude of the Fourier domain signal,
Ŷ k [r] is the magnitude of the Fourier domain signal,
Ŷ k [r−M] is the magnitude of a previous Fourier domain signal, and
â k [r] is the approximate model coefficient.
9. The system of claim 8 wherein the instructions to cause the one or more computer processors to compute the at least first order autoregressive model coefficients of the reverberation further cause the one or more computer processors to compute an autocovariance approximation of the current and previous magnitudes as follows:
a
^
k
[
r
]
=
a
^
k
(
num
)
[
r
]
a
^
k
(
den
)
[
r
]
where
â k (num) [r]=αâ k (num) [r− 1]+(1−α)( Ŷ k [r]−{circumflex over (μ)} Ŷ k [r ])( Ŷ k [r−M]−{circumflex over (μ)} Ŷ k [r ]),
â k (den) [r]=αâ k (den) [r− 1]+(1−α)( Ŷ k [r]−{circumflex over (μ)} Ŷ k [r ]) 2 ,
Ŷ k [r] is the magnitude of the Fourier domain signal,
Ŷ k [r−M] is the magnitude of a previous Fourier domain signal
{circumflex over (μ)} Ŷ k [r]=α{circumflex over (μ)} Ŷ k [r−1]+(1−α)Ŷ k [r], and
α is a parameter that determines the rate at which the estimates of the a k are allowed to adapt.