Electrical audio signal processing systems and devices – Dereverberators
Reexamination Certificate
2000-05-10
2004-05-18
Mei, Xu (Department: 2644)
Electrical audio signal processing systems and devices
Dereverberators
C381S012000, C381S094300, C379S406130, C370S290000, C708S322000
Reexamination Certificate
active
06738480
ABSTRACT:
BACKGROUND OF THE INVENTION
The present invention relates to acoustic echo cancellers which process stereophonic input signals in the frequency domain.
Acoustic echo occurs whenever there is a strong coupling between a microphone and a loudspeaker. The microphone then picks up a delayed and attenuated version of the input signal broadcast in the acoustic space by the loudspeaker. Stereophonic echo, or more generally multi-channel acoustic echo, is referred to when the microphone simultaneously picks up echoes from several loudspeakers.
Acoustic echo cancellers generally model the acoustic path between each loudspeaker and the microphone by means of an adaptive filter whose coefficients are updated by stochastic gradient algorithms: NLMS (“Normalised Least Mean Squares”), APA (“Affine Projection Algorithm”), FDAF (“Frequency Domain Adaptive Filter”), etc., or exact least squares algorithms: RLS (“Recursive Least Squares”).
It is commonly acknowledged that the performance of adaptive filtering algorithms deteriorates if multi-channel acoustic echo cancellation systems are implemented in the presence of highly correlated input signals (see F. Amand et al. “Multi-channel acoustic echo cancellation”, Proc. 4th International Workshop on Acoustic Echo and Noise Control, Roros, June 1995, pages 57-60; M. Mohan Sondhi et al., “Stereophonic Acoustic Echo Cancellation—An Overview of the Fundamental Problem”, IEEE Signal Processing Letters, Vol. 2, No. 8, August 1995, pages 148-151). Various solutions have been proposed in an attempt to overcome this problem:
using monophonic filters (A. Hirano et al., “A Compact Multi-Channel Echo Canceller with a single Adaptive Filter per Channel”, Proc. ICASSP 1992, pages 1922-1925; S. Minami, “A Stereophonic Echo Canceller Using Single Adaptative Filter”, Proc, ICASSP 1995, pages 3027-3030);
modifying time gradient algorithms (F. Amand et al., “Un algorithme d'annulation d'écho stéréo de type LMS prenant en compte l'inter-corrélation des entrées”, Fifteenth GRETSI conference, Juan-les-Pins, September 1995, pages 407-410; J, Benesty et al, “Un algorithme de projection à deux voies avec contraintes—Application à l'annulation d'écho acoustique stéréophonique”, Fifteenth GRETSI conference, Juan-les-Pins, September 1995, pages 387-390);
de-correlating signals before broadcasting them (J. Benesty et al., “A Hybrid Mono/Stereo Acoustic Echo Canceller”, IEEE Workshop on application of signal processing and acoustics (WASPAA'97)).
The echo cancellers according to the invention find applications in multi-channel communication systems in particular in video-conferencing systems (see P. Heitkamper et al., “Stereophonic and multichannel Hands-Free Speaking”, Proc. 4th International Workshop on Acoustic Echo and Noise Control, Roros, June 1995, pages 53-56; Y. Mahieux et al., “Annulation d'écho en téléconférence stéréophonique”, 14th GRETSI, Juan-les-Pins, September 1993, pages 515-518), in hands-free telephones and in speech recognition systems (see M. Glanz et al., “Speech Recognition In Cars With Noise Suppression and Car Radio Compensation”, 22nd ISATA, Florence, May 1990, pages 509-516; F. Berthault et al., “Stereophonic Acoustic Echo Cancellation—Application to speech recognition: Some experimental results”, 5th International Workshop on Acoustic Echo and Noise Control, London, September 1997, pages 96-99).
Frequency domain stereophonic echo cancellers implement a method wherein first and second input signals (x
1
, x
2
) are applied to an echo generator system and an observation signal (z) is picked up at an output of said system, the input signals being digitally sampled and processed in successive blocks of 2N samples with frequency domain transformation according to a set of 2N frequencies. In accordance with this method, the processing of a block of 2N samples comprises the steps of:
transforming the first input signal from the time domain to the frequency domain to obtain a vector
X
1
having 2N complex components relating to the set of 2N frequencies, including spectral components of the first input signal relating to a sub-set of the set of 2N frequencies;
transforming the second input signal from the time domain to the frequency domain to obtain a vector
X
2
having 2N complex components relating to the set of 2N frequencies, including spectral components of the second input signal relating to said sub-set of frequencies;
multiplying term by term the vector
X
1
by a vector
H
1
of 2N complex coefficients to produce first estimated spectral echo components relating to the frequencies of the sub-set;
multiplying term by term the vector
X
2
by a vector
H
2
of 2N complex coefficients to produce second estimated spectral echo components relating to the frequencies of the sub-set;
adding the first and second estimated spectral echo components relating to each frequency of the sub-set to obtain a spectral component belonging to a vector of 2N estimated spectral total echo components;
transforming the vector of 2N estimated spectral total echo components from the frequency domain to the time domain to obtain an estimated total echo;
subtracting the estimated total echo from the observation signal to produce an error signal;
transforming the error signal from the time domain to the frequency domain to obtain a vector
E
of 2N spectral components of the error signal relating to the set of 2N frequencies; and
updating the vectors
H
1
and
H
2
for the processing of the next block, on the basis of the vectors
X
1
,
X
2
and
E
.
In the known systems, said sub-set of frequencies represents the entire set of 2N frequencies.
Usually (stereophonic FDAF algorithm), the updating of the vectors
H
1
and
H
2
for the processing of the next block takes account of the energy gradient of the error signal, estimated by
∇
i
=
X
*
i
{circle around (X)}
E
for the vector
H
i
(i=1 or 2), where {circle around (X)} denotes the term-by-term product of two vectors and (*) denotes complex conjugation.
The gradient is generally normalised:
∇
Ni
=
B
i
{circle around (X)}
∇
i
for i=1 or 2, where
B
i
is a vector of size 2N, whose term corresponding to a frequency f is the inverse of the spectral energy P
ii
(f) of the i-th input signal evaluated at the frequency f (in other words, P
ii
(f)=<X
i
(f).X
i
(f)*> is a current average of |X
i
(f)|
2
=X
i
(f).X
i
(f)*, where X
i
(f) is the component of the vector
X
i
relating to the frequency f).
In addition, a constraint is often placed on the normalised gradient in order to retain only the linear convolution terms in the frequency calculation of the gradients:
∇
Ci
=
C
.
∇
Ni
for i=1 or 2, where
C
denotes a constant constraint matrix.
The echo estimation filters are finally adapted by
H
i
(k+1)=
H
i
(k)+&mgr;.
∇
Ci
(k) for i=1 or 2, the index k numbering the successive analysis blocks. The coefficient &mgr;, lying between 0 and 1, is the adaptation step.
It is noted that each processing channel is subjected to a separate adaptation determined by the error signal and the input signal relating to this channel. This explains the identification errors which might be made by the algorithm in the presence of correlated input signals: two estimation errors for the vectors
H
1
and
H
2
can compensate for one another in the error signal while the algorithm is unable to correct them.
An object of the present invention is to propose another method of adapting stereophonic frequency filters which allows a certain degree of correlation between the signals to be taken into account.
SUMMARY OF THE INVENTION
Accordingly, the invention proposes a method as outlined above, further comprising computing a spectral energy P
11
(f) of the first input signal, a spectral energy P
22
(f) of the second input signal, an inter-spectral energy P
12
(f) of the first and second input signals and a coherence value &Ggr;(f) for each of the frequencies f in the set of 2N f
Berthault Frederic
Capman François
Matra Nortel Communications
Mei Xu
Top, Pruner & Hu, P.C.
Tran Con P.
LandOfFree
Method and device for cancelling stereophonic echo with... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method and device for cancelling stereophonic echo with..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and device for cancelling stereophonic echo with... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3256686