Adaptive filtering with flexible selection of algorithm complexity and performance
View Patent ↗An adaptive filter configured to use multiple algorithm species that differ in the quality of echo suppression and respective burdens imposed on the computational resources of the host communication device. Depending on the available computational budget, the adaptive filter selects an algorithm species that, while supporting a relatively high quality of echo suppression, involves a relatively low risk of overwhelming the computational resources. The adaptive filter monitors changes in the available computational budget and, if appropriate or necessary, can change the algorithm species to maintain a quality of echo suppression that is optimal for the current computational budget. If a change of the algorithm species is initiated, then at least a portion of internal algorithm data from the previously running algorithm species might be transferred for use in the subsequent algorithm species.
1. A device, comprising:
an algorithm module having a plurality of algorithm species, each adapted to suppress echo in a communication signal; and
a selector module operatively coupled to the algorithm module and adapted to:
select a first algorithm species from said plurality based on an available computational budget;
configure the algorithm module to run the first algorithm species to perform said echo suppression;
detect a change in the available computational budget;
select a second algorithm species from said plurality based on the detected change in the available computational budget; and
reconfigure the algorithm module to run the second, instead of the first, algorithm species to perform said echo suppression, wherein:
the first algorithm species has a first projection order;
the second algorithm species has a second projection order different from the first projection order by a value of one, with both the first projection order and the second projection order being positive integers; and
wherein the algorithm module is adapted to transfer at least a portion of internal algorithm data generated by the first algorithm species for use in the run of the second algorithm species.
2. The device of claim 1 , wherein the internal algorithm data comprises at least one of: an autocorrelation vector or matrix, a forward-prediction-coefficients vector, a de-correlation filter, and a backward-prediction-coefficients vector.
3. The device of claim 1 , wherein the plurality of algorithm species comprises at least two algorithm species whose projection orders differ by a value that is greater than one.
4. The device of claim 3 , wherein a number of taps is fixed within said plurality of algorithm species.
5. The device invention of claim 1 , wherein:
the plurality of algorithm species comprises a normalized least mean square (NLMS) algorithm having a projection order of one;
the selector module is further adapted to reconfigure the algorithm module to run the NLMS algorithm instead of the second algorithm species to perform said echo suppression, without running any other algorithm species of the plurality between the second algorithm species and the NLMS algorithm; and
the second projection order is greater than two.
6. The device of claim 1 , wherein:
the second projection order is smaller than the first projection order; and
to perform the transfer, the algorithm module is configured to perform one or more of the following:
shrink an autocorrelation matrix generated by the first algorithm species by removing a row and a column from said autocorrelation matrix;
reduce a length of an error vector generated by the first algorithm species by removing an element from said error vector;
reduce a length of a forward-prediction-coefficients vector generated by the first algorithm species by removing an element from said forward-prediction-coefficients vector;
reduce a length of a backward-prediction-coefficients vector generated by the first algorithm species by removing an element from said backward-prediction-coefficients vector;
reduce a length of a de-correlated error vector generated by the first algorithm species by removing an element from said de-correlated error vector; and
reduce a length of a de-correlation filter generated by the first algorithm species by removing an element from said de-correlation filter.
7. The device of claim 1 , wherein:
the second projection order is greater than the first projection order; and
to perform the transfer, the algorithm module is configured to perform one or more of the following:
enlarge an autocorrelation matrix generated by the first algorithm species by calculating missing autocorrelation values for an additional row and an additional column;
increase a length of an error vector generated by the first algorithm species by padding said error vector with a zero element;
increase a length of a forward-prediction-coefficients vector generated by the first algorithm species by padding said forward-prediction-coefficients vector with a zero element;
increase a length of a backward-prediction-coefficients vector generated by the first algorithm species by padding said backward-prediction-coefficients vector with a zero element;
increase a length of a de-correlated error vector generated by the first algorithm species by padding said de-correlated error vector with a zero element; and
increase a length of a de-correlation filter generated by the first algorithm species by padding said de-correlation filter with a zero element.
8. A method of adaptive filtering, comprising the steps of:
based on an available computational budget, selecting a first algorithm species from a plurality of algorithm species, each adapted to suppress echo in a communication signal of a communication device;
configuring the communication device to run the first algorithm species to perform said echo suppression;
detecting a change in the available computational budget;
selecting a second algorithm species from said plurality based on the detected change in the available computational budget; and
reconfiguring the communication device to run the second, instead of the first, algorithm species to perform said echo suppression, wherein:
the first algorithm species has a first projection order;
the second algorithm species has a second projection order different from the first projection order by a value of one, with both the first projection order and the second projection order being positive integers; and
the step of reconfiguring comprises transferring at least a portion of internal algorithm data generated by the first algorithm species for use in running the second algorithm species.
9. The method of claim 8 , wherein the internal algorithm data comprises at least one of: an autocorrelation vector or matrix, a forward-prediction-coefficients vector, a de-correlation filter, and a backward-prediction-coefficients vector.
10. The method of claim 8 , wherein the plurality of algorithm species comprises at least two algorithm species whose projection orders differ by a value that is greater than one.
11. The method of claim 8 , wherein:
the second projection order is smaller than the first projection order; and
the step of transferring comprises one or more of:
shrinking an autocorrelation matrix generated by the first algorithm species by removing a row and a column from said autocorrelation matrix;
reducing a length of an error vector generated by the first algorithm species by removing an element from said error vector;
reducing a length of a forward-prediction-coefficients vector generated by the first algorithm species by removing an element from said forward-prediction-coefficients vector;
reducing a length of a backward-prediction-coefficients vector generated by the first algorithm species by removing an element from said backward-prediction-coefficients vector;
reducing a length of a de-correlated error vector generated by the first algorithm species by removing an element from said de-correlated error vector; and
reducing a length of a de-correlation filter generated by the first algorithm species by removing an element from said de-correlation filter.
12. The method of claim 8 , wherein:
the second projection order is greater than the first projection order; and
the step of transferring comprises one or more of:
enlarging an autocorrelation matrix generated by the first algorithm species by calculating missing autocorrelation values for an additional row and an additional column;
increasing a length of an error vector generated by the first algorithm species by padding said error vector with a zero element;
increasing a length of a forward-prediction-coefficients vector generated by the first algorithm species by padding said forward-prediction-coefficients vector with a zero element;
increasing a length of a backward-prediction-coefficients vector generated by the first algorithm species by padding said backward-prediction-coefficients vector with a zero element;
increasing a length of a de-correlated error vector generated by the first algorithm species by padding said de-correlated error vector with a zero element; and
increasing a length of a de-correlation filter generated by the first algorithm species by padding said de-correlation filter with a zero element.