Optimal ratio estimator for multisensor systems
View Patent ↗A signal processing technique can be effectively used for source separation, signal enhancement, and noise reduction when using a twin microphone system. The class of stochastic signals for which ratio-estimates can be computed from histograms is defined. This class fits real-world signals of interest such as voice signals. Theoretical computation in closed form of the optimal estimator for this class of signals is disclosed. Two practical implementation solutions are disclosed, as is a practical solution to exploit an echoic environment model. Furthermore, two novel techniques for signal demixing are presented. The application of the optimal estimator and the suboptimal estimator to the case of more than two channels is disclosed.
1. Apparatus for ratio estimation comprising:
first and second microphone transducers for providing respective signal data from first and second mobile phone microphone transducers;
means for transforming said respective signal data into respective frequency domain representations;
means for computing instantaneous ratios between said respective frequency domain representations;
means for creating respective histograms of said instantaneous ratios;
means for finding and updating peaks and permutations in said respective histograms;
means for computing weight data for individual frequency components of said respective frequency domain representations by utilizing said peaks and permutations in said respective histograms;
means for optionally filtering said weight data to derive optionally filtered weight data;
means for transforming said optionally filtered weight data into a time-domain representation; and
means for outputting said time-domain representation.
2. Apparatus for ratio estimation comprising:
first and second microphone transducers for providing respective signal data from said first and second microphone transducers;
means for transforming said respective signal data into respective frequency domain representations;
means for computing instantaneous ratios between said respective frequency domain representations;
means for creating respective histograms of said instantaneous ratios;
means for finding and updating peaks and permutations in said respective histograms;
means for computing weight data for individual frequency components of said respective frequency domain representations by utilizing said peaks and permutations in said respective histograms;
means for optionally filtering said weight data to derive optionally filtered weight data;
means for transforming said optionally filtered weight data into a time-domain representation; and
means for outputting said time-domain representation to a respective mobile phone transducer.
3. A method of modeling echoic environments, comprising:
computing SDFT ratios R({tilde over (ω)},k);
estimating {circumflex over (R)}({tilde over (ω)},k) for each histogram peak;
determining permutation for assembling R-estimates together;
identify parameters of the environment model for R({tilde over (ω)},k);
recomputing R-estimates based on the parameters; and
recovering independent signals by signal demixing.
4. The method of claim 3 , wherein determining permutation for assembling R-estimates together is given by
π
opt
=
argmax
π
∑
j
-
1
n
{
k
,
X
1
(
ϖ
,
k
)
X
2
(
ϖ
,
k
)
-
R
π
(
j
)
(
ϖ
)
<
∂
and
X
1
(
ϖ
ref
,
k
)
X
2
(
ϖ
ref
,
k
)
-
R
j
(
ϖ
ref
)
<
∂
}
.
5. The method of claim 3 , wherein recovering independent signals by signal demixing comprises:
partitioning of a complex plan by a Voronoi tessellation on the set of points R 1,2, . . . ,n ;
spectral weighting of mixtures in the frequency domain, with the characteristic function of each Voronoi set; and
inverting of STFT signals obtained by spectral weighting.
6. A method of ratio estimation comprising the steps of:
deriving respective signal data from two sources;
transforming said respective signal data into respective frequency domain x 1 (t)→X 1 ({tilde over (ω)}),x 2 (t)→X 2 ({tilde over (ω)});
computing the instantaneous ratios
X
2
(
ϖ
)
X
1
(
ϖ
)
between said respective frequency domain representations;
computing the weights for individual frequency components of said respective frequency domain representations: {tilde over (H)} k (ω) or {tilde over (H)} 1k , {tilde over (H)} 2k ;
optionally filtering said weight data to derive optionally filtered weight data;
transforming said optionally filtered weight data into a time-domain representation: Ŝ k ({tilde over (ω)})→ŝ k (t); and
outputting said time-domain representation.
7. The method of claim 6 , wherein optionally filtering said weight data comprises filtering data using the direct method: Ŝ k ={tilde over (H)} k X 1 .
8. The method of claim 6 , wherein optionally filtering said weight data comprises filtering data using the Wiener filtering method:
Ŝ k ={tilde over (H)} 1k X 1 +{tilde over (H)} 2k X 2
where
{tilde over (H)} 1k =(ν 1 +ν 2 +ν 3 )( {overscore (R)} 1 ν 1 ∂ 1k
+{overscore (R)} 2 ν 2 ∂ 2k +{overscore (R)} 3 ν 3 ∂ 3k )−
( {overscore (R)} 1 ν 1 +{overscore (R)} 2 ν 2 +{overscore (R)} 3 ν 3 )
(ν 1 ∂ 1k +ν 2 ∂ 2k +ν 3 ∂ 3k )
and
{tilde over (H)} 2k =−( R 1 ν 1 +R 2 ν 2 +R 3 ν 3 )
( {overscore (R)} 1 ν 1 ∂ 1k +{overscore (R)} 2 ν 2 ∂ 2k +{overscore (R)} 3 ν 3 ∂ 3k )+
( R 1 {overscore (R)} 1 ν 1 +R 2 {overscore (R)} 2 ν 2 +R 3 {overscore (R)} 3 ν 3 )
(ν 1 ∂ 1k +ν 2 ∂ 2k +ν 3 ∂ 3k ).