Neural-AFC: Learning-Based Step-Size Control for Adaptive Feedback Cancellation with Closed-Loop Model Training

Acoustic feedback arises from the leakage of sound from the loudspeaker back to the microphone and is a common problem in hearing aids and public address systems. One popular solution to this problem is adaptive feedback cancellation (AFC) which utilizes an adaptive filter (AF) to estimate and subse...

Full description

Saved in:
Bibliographic Details
Main Authors: Soleimani, Behrad, Schepker, Henning, Mirbagheri, Majid
Format: Conference Proceeding
Language:English
Subjects:
Adaptation models
Adaptive Filter
Adaptive filters
Correlation
Deep Neural Network
Feedback Cancellation
Hearing Aid
Hearing aids
Loudspeakers
Public Address Systems
Recurrent neural networks
Training
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Acoustic feedback arises from the leakage of sound from the loudspeaker back to the microphone and is a common problem in hearing aids and public address systems. One popular solution to this problem is adaptive feedback cancellation (AFC) which utilizes an adaptive filter (AF) to estimate and subsequently subtract the feedback component from the microphone. Due to the closed-loop system the loudspeaker signal and the incoming signal exhibit a high correlation. However, model-based and more recently learning-based AF methods typically neglect this correlation in their derivation, leading to sub-optimal performance in closed-loop scenarios. In this paper, we propose Neural-AFC, a recurrent neural network (RNN) designed for AFC step-size control optimization, that addresses this problem by including not only the adaptive filter but also the acoustic feedback path and the system gain in the recurrence model. Our experiment results show that when trained within this closed-loop model, Neural-AFC improves the steady-state performance and (re)convergence rate compared to conventional and open-loop RNN-based AF in both hearing aid and public address system applications.
ISSN:2379-190X
DOI:10.1109/ICASSP49357.2023.10094841