Geometric source separation signal processing technique
AssigneeThomson Licensing
InventorsLucas Cristobal Parra, Christopher Vincent Alvino, Clay Douglas Spence, Craig Langdale Fancourt
Life of this patent
Estimating the term…
Abstract
A computer system that processes mixtures of signals, such as speech and noise sources derived from multiple simultaneous microphone recordings, in order to separate them into their underlying sources. A source separation routine optimizes a filter structure by minimizing cross powers of multiple output channels while enforcing geometric constraints on the filter response. The geometric constraints enforce desired responses for given locations of the underlying sources, based on the assumption that the sources are localized in space.
Claims15
An apparatus for separating a convolutively mixed signal into a plurality of source signals comprising:
- a plurality of signal sensors, each sensor residing at a sensor location for sensing a source signal of the plurality of source signals at the sensor location and providing a sensor signal representing the sensed source signal;
- a multiple input, multiple output signal filter for filtering a convolutively mixed signal, said convolutively mixed signal comprising sensor signals from each of said plurality of signal sensors;
- means for estimating cross powers of multiple output channels, said cross power being estimated for each of the multiple output channels at multiple times;
- means for computing a set of source locations r of the plurality of source signals for use in constraining subsequent source separation, each source location of the set of source locations r being defined by spatial relation to the sensor locations;
- means for adapting the multiple input, multiple output signal filter to minimize the cross powers while enforcing a desired filter response for the set of source locations;
- and
- said multiple input, multiple output signal filter filtering the convolutively mixed signal in order to separate the convolutively mixed signal into said plurality of source signals, wherein said filtering utilizes said computed set of source locations r as a constraint in determining separation of the plurality of source signals.
A method of separating a composite signal into a plurality of component signals comprising the steps of computing a cross power spectra Ryy(t,ω) of the plurality of component signals, simultaneously diagonalizing Ryy(t,ω) with an algorithm defined as:
- J
-
- (
- W
- )
- =
- ∑
- t
- ,
- ω
-
-
- α
-
- (
- ω
- )
-
-
- R
- yy
-
- (
- t
- ,
- ω
- )
- -
- diag
-
- (
- R
- yy
-
- (
- t
- ,
- ω
- )
- )
-
- F
- 2
- with the factorization Ryy(t,ω)≈W(ω)Rxx(t,ω)WH(ω), applying a linear constraint to the algorithm to create filter coefficients, and filtering the composite signal based on the filter coefficients to separate said composite signal into a plurality of component signals, further comprising the step of choosing a linear constraint from a set of two linear constraints given by a first linear constraint
- diag(W(ω)D(ω, r))=I
- and a second linear constraint
- (W(ω)D(ω, r))=I
- wherein the above steps are performed using a computing system.
A method of separating a composite signal into a plurality of component signals comprising the steps of computing a cross power spectra Ryy(t,ω) of the plurality of component signals, simultaneously diagonalizing Ryy(t,ω) with an algorithm defined as:
- J
-
- (
- W
- )
- =
- ∑
- t
- ,
- ω
-
-
- α
-
- (
- ω
- )
-
-
- R
- yy
-
- (
- t
- ,
- ω
- )
- -
- diag
-
- (
- R
- yy
-
- (
- t
- ,
- ω
- )
- )
-
- F
- 2
- with the factorization Ryy(t,ω)≈W(ω)Rxx(t,ω)WH(ω), applying a linear constraint to the algorithm to create filter coefficients, and filtering the composite signal based on the filter coefficients to separate said composite signal into a plurality of component signals and applying a regularization term to the diagonalizing algorithm of
- J(ω)=∥diag(W(ω)D(ω,r))−I∥F2
- wherein the above steps are performed using a computing system.
Dependencies are inferred from the claim text.