Learnable Kalman Filtering for Nonlinear Acoustic Echo Paths
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional AEC algorithms struggle with modeling nonlinearity and require tuning of control parameters, leading to limitations in fast convergence and effectiveness in double-talk scenarios, while deep learning-based methods face challenges with continuously changing echo paths.
Innovation Solution
A hybrid method combining a frequency domain Kalman filter (FDKF) with a deep neural network (DNN) is employed, where the DNN estimates nonlinear distortions and transition factors to enhance the FDKF algorithm, leveraging both deep learning and adaptive filtering algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional adaptive filtering algorithms (NLMS, RLS) are used for AEC, then echo removal can be achieved through linear transfer function estimation, but nonlinearity modeling is missing and control parameters need to be tuned
Solution Approach 1:
The patent replaces conventional DSP-based adaptive filtering algorithms with a deep learning-based approach. The neural network automatically learns non-linear echo path characteristics from training data, eliminating the need for manual parameter tuning while maintaining robust echo removal capability across varying acoustic conditions.
Solution Approach 2:
The patent transforms the echo cancellation approach by changing from fixed linear models to adaptive non-linear models. The neural network dynamically adjusts its internal parameters based on training data, enabling it to model complex non-linear echo paths without requiring external parameter tuning.
2Reliability
If deep learning-based methods are used for AEC, then nonlinearity can be modeled, but the method fails to handle continuously changing echo paths effectively
Solution Approach 1:
The patent employs a dynamic echo path model that adapts to continuously changing acoustic conditions. The neural network processes time-varying input signals and updates its internal representations in real-time, enabling it to track and compensate for changing echo paths while maintaining non-linearity modeling capability.
Solution Approach 2:
The patent incorporates feedback mechanisms where the neural network continuously monitors the acoustic environment and adjusts its processing accordingly. This feedback loop enables the system to adapt to continuously changing echo paths while maintaining accurate non-linear modeling through iterative learning and adjustment.
3Reliability
If frequency domain Kalman filter is used for AEC, then robustness in double-talk scenarios is achieved, but convergence rate is slower compared to other methods
Solution Approach 1:
The patent replaces the traditional Kalman filter with a deep learning-based approach that combines the robustness of adaptive filtering with the convergence speed of neural networks. The neural network learns optimal filtering parameters from training data, achieving both fast convergence and robustness in double-talk scenarios without requiring manual parameter tuning.
Data Source
AI summary
A method and apparatus comprising computer code configured to cause a processor or processors to receive an audio signal obtained from a microphone, input the audio signal into a neural-network based AEC model, and output an AEC signal from the neural-network based AEC model in which AEC is applied to the audio signal, and the AEC signal is a version of the audio signal in which acoustic echo noise of the audio signal is suppressed and target audio of the audio signal is sustained.