Subband Acoustic Echo Cancellation With Diagonal Regularization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current acoustic echo cancellation techniques for spatial audio communication and teleconferencing on mobile devices face challenges due to high computational complexity and instability, particularly in handling multi-channel and stereo playback scenarios with long echo impulse responses.
Innovation Solution
Implementing subband adaptive filters using diagonally regularized correlation matrices to iteratively determine gain vectors and update filter coefficients, reducing echoes by processing subband signal sequences, and employing a technique to diagonally regularize the inverse correlation matrix to ensure robustness and computational efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If time-domain filter implementations are used for acoustic echo cancellation, then the filtering operation is straightforward, but the computational complexity becomes very high due to long echo impulse responses requiring thousands of taps
Solution Approach 1:
The patent divides the frequency spectrum into multiple subbands and processes each subband separately using shorter filters. This segmentation approach transforms a single complex time-domain filter with thousands of taps into multiple simpler filters operating in parallel on divided frequency segments, thereby reducing overall computational complexity while maintaining echo cancellation effectiveness
Solution Approach 2:
The patent transitions from time-domain filtering to frequency-domain filtering by applying Fourier transforms. This dimensional change allows the system to exploit the convolution theorem, where time-domain convolution becomes frequency-domain multiplication, significantly reducing the computational burden from O(N²) to O(N log N) complexity
2Reliability
If multiple adaptive filters are applied in parallel to every frequency bin, then the acoustic echo cancellation effectiveness is improved, but the CPU usage and computational load increase significantly
Solution Approach 1:
The patent segments the frequency spectrum into multiple subbands and applies adaptive filtering only within each subband rather than across the entire frequency range. This segmentation reduces the number of adaptive filters needed from one per frequency bin to one per subband, significantly reducing CPU usage while maintaining cancellation effectiveness through localized adaptation
Solution Approach 2:
The patent applies adaptive filtering selectively to specific subbands rather than uniformly to all frequency bins. By concentrating computational resources on subbands where echo cancellation is most needed and using shorter filter lengths, the system achieves sufficient cancellation effectiveness with reduced overall computational load
3Productivity
If diagonal regularization of the correlation matrix is applied, then the computational efficiency and stability are improved, but the filtering accuracy may be reduced compared to full matrix inversion
Solution Approach 1:
The patent modifies the correlation matrix by applying diagonal regularization, which changes the matrix parameters from full covariance to diagonally-loaded form. This parameter change adds a scaled identity matrix to the correlation matrix, ensuring numerical stability and enabling efficient inversion through closed-form solutions, while the scaling factor is optimized to maintain filtering accuracy
Solution Approach 2:
The patent applies different processing strategies to different parts of the correlation matrix structure. By focusing on the diagonal elements and applying regularization specifically there, the system achieves computational efficiency without significantly degrading the overall filtering accuracy, as the off-diagonal elements (cross-correlations) are handled through the subband decomposition
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
Devices, methods and computer programs for acoustic echo cancellation (AEC) based on one or more diagonally regularized correlation matrices are disclosed. At least some example embodiments may allow a smart and efficient solution for multi-channel and/or stereo AEC, providing improved performance and ease of implementation.