RNN Acoustic Echo Suppression for Non-Linear Echo and Noise
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing acoustic echo cancellation (AEC) methods are inefficient for non-linear echo distortion and non-stationary noises, particularly in full-duplex speech communication and speech recognition, as they fail to effectively handle changes in the echo path and environmental noise.
Innovation Solution
A NeuralEcho model using a recurrent neural network (RNN) with automatic gain control (AGC) is trained in a unified model to perform acoustic echo cancellation and noise suppression, leveraging second-order statistics and multi-head self-attention for dynamic feature emphasis, and integrating feature-wise linear modulation to separate target speech from echo and interference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If adaptive filtering methods are used for echo cancellation, then the system complexity is low and robustness is good, but the method is inefficient for non-linear echo distortion and non-stationary noises
Solution Approach 1:
The patent combines adaptive filtering methods with deep neural networks to form a hybrid system. The adaptive filter handles linear echo cancellation while the DNN component processes non-linear echo distortion and non-stationary noises, merging the strengths of both approaches to resolve the contradiction between robustness and adaptability.
Solution Approach 2:
The hybrid system is designed to perform multiple functions: linear echo cancellation via adaptive filtering, non-linear echo distortion suppression via DNN, and non-stationary noise handling via the same DNN component. This multi-functional design allows the system to maintain robustness while gaining versatility in handling various echo and noise conditions.
2Adaptability or versatility
If deep neural networks are employed for echo suppression, then non-linear echo distortions and echo path delay are better handled, but the system complexity increases
Solution Approach 1:
The patent segments the echo cancellation task into two parts: linear echo cancellation handled by adaptive filtering and non-linear echo distortion handled by deep neural networks. This segmentation allows each component to specialize in specific aspects, improving overall performance while managing system complexity through division of labor.
Solution Approach 2:
The hybrid system applies partial action by using adaptive filtering for the linear component and DNN for the non-linear component, rather than using a complete DNN solution for all aspects. This approach achieves the necessary adaptability for non-linear distortions without the full complexity burden of a pure deep learning system.
3Reliability
If a unified model with AGC is used for AEC and noise suppression, then echo and noise reduction is improved, but the model complexity increases
Solution Approach 1:
The patent merges acoustic echo cancellation and noise suppression into a unified model that processes both tasks simultaneously. The model integrates AEC functionality with noise suppression capabilities and AGC, creating a single integrated system that improves overall echo and noise reduction performance while managing complexity through unified architecture.
Solution Approach 2:
The unified model performs multiple functions including echo cancellation, noise suppression, and automatic gain control within a single architecture. This multi-functional design improves reliability by handling multiple audio enhancement tasks concurrently while avoiding the complexity of separate dedicated systems for each function.
Data Source
AI summary
A method of acoustic echo suppression using a recurrent neural network, performed by at least one processor, is provided. The method includes receiving a microphone signal and a far-end reference signal, estimating an echo suppressed signal and an echo signal based on the microphone signal and the far-end reference signal, estimating enhancement filters for the microphone signal based on the echo suppressed signal and the echo signal, generating an enhanced signal based on the enhancement filters, and adjusting the enhanced signal using automatic gain control (AGC) and outputting the adjusted signal.