Stochastic Vector Echo Cancellation for Convergence Speed
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current network echo cancellation techniques face challenges in balancing convergence speed and quality with computational complexity, particularly due to slow convergence rates of linear adaptive filtering methods like NLMS and inefficiencies in partial-update techniques such as SPU+PNLMS, which are numerically intensive and resource-consuming.
Innovation Solution
A stochastic vector-based approach is introduced, where a first circuit generates samples by filtering an input vector using a filter vector and a stochastic vector with random components, and a second circuit updates a subset of filter taps based on the stochastic vector's random values, reducing computational complexity and improving convergence speed while maintaining quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If NLMS technique is used for adaptive filtering, then computational complexity is reduced, but convergence rate becomes slow
Solution Approach 1:
The filter coefficients are divided into multiple segments or groups, allowing selective updates of only those segments that require adjustment. This segmentation enables the system to maintain lower computational complexity while improving convergence rate by focusing updates on relevant portions of the filter rather than updating all coefficients uniformly.
Solution Approach 2:
Instead of updating all filter coefficients at every iteration, the invention updates only a partial subset of coefficients selected based on stochastic criteria. This partial action reduces the computational burden per iteration while still achieving convergence, effectively trading off complete updates for faster overall convergence with lower complexity.
2Speed
If PNLMS technique is used for adaptive filtering, then convergence rate is improved, but computational complexity increases
Solution Approach 1:
The invention applies partial updates to filter coefficients by selecting only a subset for update based on stochastic selection rather than updating all coefficients as in PNLMS. This reduces the computational complexity from O(N) per iteration to approximately O(N/M) where M is the number of segments, while maintaining improved convergence characteristics.
Solution Approach 2:
Different segments of the filter coefficients are treated differently, with only selected segments updated at each iteration based on stochastic selection. This local quality approach allows the system to apply computational effort selectively to portions of the filter that benefit most from updates, reducing overall complexity while maintaining convergence performance.
3Speed
If multi-segmental filter coefficient approaches are used, then convergence speed is increased, but additional computations for window segment boundaries are required
Solution Approach 1:
The invention changes the parameter selection criterion from deterministic window boundary positions to stochastic random selection. This parameter change eliminates the need for complex window segment boundary computations while maintaining the benefits of segmented filter updates, as segments are selected randomly rather than based on fixed boundary positions.
Solution Approach 2:
Instead of dividing the filter into fixed segments and updating only those at boundaries, the invention inverts the approach by randomly selecting which segments to update. This inversion eliminates the need for boundary detection and window management computations, simplifying the implementation while maintaining convergence speed improvements.
Data Source
AI summary
An apparatus generally having a first circuit and a second circuit is disclosed. The first circuit may be configured to generate a first sample by filtering an input vector based on (a) a filter vector and (b) a stochastic vector. Each of a plurality of components in the stochastic vector generally has a respective random value. The first circuit may also be configured to generate a second sample as a difference between a third sample and the first sample. The third sample may be received from a network as an echo. The second circuit may be configured to update a subset of a plurality of taps of the filtering where a corresponding one of the components of the stochastic vector has a first value of the random values.