Deep Adaptive Acoustic Echo Cancellation Using Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional acoustic echo cancellation (AEC) processing struggles to fully remove nonlinear and time-varying echo signals due to mechanical noise and device movement, often resulting in residual echo, especially when linear methods are used.
Innovation Solution
The implementation of deep adaptive AEC processing, which integrates a deep neural network (DNN) with linear adaptive filtering to generate nonlinear reference signals and step-size data, allowing for improved echo cancellation by freezing or increasing adaptation based on the presence of local speech, thereby accurately modeling and removing echo signals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If linear adaptive filtering is used for echo cancellation, then the system complexity is low, but the echo cancellation performance deteriorates due to inability to handle nonlinear and time-varying echo signals
Solution Approach 1:
The patent combines linear adaptive filtering with deep neural network processing to create a hybrid echo cancellation system. The DNN processes nonlinear and time-varying components while the linear filter handles stationary components, achieving superior cancellation performance without excessive complexity increase.
Solution Approach 2:
The system dynamically switches between different processing modes based on echo characteristics. When echo paths are time-varying or nonlinear, the DNN component is activated to adapt to changing conditions, while linear filtering handles stable conditions, optimizing performance across varying scenarios.
2Reliability
If deep neural network is integrated with linear adaptive filtering, then the echo cancellation performance improves, but the device complexity increases
Solution Approach 1:
The echo cancellation system is segmented into distinct functional components: a DNN module for nonlinear/time-varying echo processing and a linear adaptive filter for stationary echo processing. This segmentation allows each component to be optimized independently and reduces overall system complexity through modular design.
Solution Approach 2:
The patent introduces an intermediary processing stage where the DNN generates nonlinear reference signals that are then fed into the linear adaptive filtering stage. This intermediary approach allows the complex nonlinear processing to be separated from the simpler linear filtering, managing complexity while maintaining performance.
3Speed
If the adaptive filter continuously adapts, then the tracking of time-varying echo paths improves, but the convergence to steady-state increases loss of time
Solution Approach 1:
The adaptive filter's behavior is dynamically adjusted based on the time-varying characteristics of the echo path. During transient periods when echo paths change, the adaptation rate is increased for fast tracking. During steady-state periods, adaptation is reduced or frozen to prevent unnecessary adjustments and accelerate convergence.
Solution Approach 2:
The system uses feedback mechanisms to monitor echo path stability and dynamically control the adaptation process. When stability is detected, the feedback signal triggers a reduction in adaptation activity, allowing the system to converge faster to steady-state without sacrificing tracking capability during transitions.
Data Source
AI summary
A system configured to perform deep adaptive acoustic echo cancellation (AEC) to improve audio processing. Due to mechanical noise and continuous echo path changes caused by movement of a device, echo signals are nonlinear and time-varying and not fully canceled by linear AEC processing alone. To improve echo cancellation, deep adaptive AEC processing integrates a deep neural network (DNN) and linear adaptive filtering to perform echo and/or noise removal. The DNN is configured to generate a nonlinear reference signal and step-size data, which the linear adaptive filtering uses to generate output audio data representing local speech. The DNN may generate the nonlinear reference signal by generating mask data that is applied to a microphone signal, such that the reference signal corresponds to a portion of the microphone signal that does not include near-end speech.