Hearing Aid Feedback Control Using ML Feedback Path Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ML-based acoustic feedback cancellation systems in hearing aids face challenges in balancing convergence rate and steady-state error, leading to compromised sound quality and speech intelligibility, with conventional techniques struggling to accurately estimate and react to feedback path changes.
Innovation Solution
A method for training a ML model using a convolutional layer, followed by a fully connected layer and a long-short term memory layer, to estimate an intermediate signal, which is then post-processed to improve the feedback-free signal, allowing for faster convergence and lower steady-state errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If state-of-the-art adaptive filtering algorithms (e.g., variable step-size algorithms) are used for acoustic feedback cancellation, then convergence rate may be improved, but steady-state error increases leading to inaccurate feedback estimation and failure to react quickly to feedback path changes
Solution Approach 1:
The patent introduces an intermediary signal (feedback path impulse response) as a intermediate representation between the raw feedback signal and the final cancellation control. By estimating this intermediate impulse response first, the system can then derive accurate feedback estimates through convolution operations, resolving the trade-off between convergence speed and estimation accuracy.
Solution Approach 2:
The patent replaces traditional mechanical adaptive filtering algorithms with a machine learning-based system that uses neural networks to estimate the feedback path impulse response. This substitution enables faster convergence while maintaining high estimation accuracy through the learned representations, overcoming the limitations of conventional algorithms.
2Speed
If ML-based AFC systems are configured to directly predict the feedback-free signal, then convergence may be faster, but sound quality and speech intelligibility are compromised due to artifacts in the predicted signal
Solution Approach 1:
Instead of directly predicting the feedback-free signal, the system uses an intermediary approach by first estimating the feedback path impulse response. This intermediate representation serves as a mediator that can be post-processed to remove artifacts, thereby eliminating the harmful effects while maintaining fast convergence benefits of direct ML prediction.
Solution Approach 2:
The system performs preliminary estimation of the feedback path impulse response before generating the final feedback cancellation signal. This preliminary action allows for post-processing steps to be applied, removing artifacts from the intermediate estimate before it is used for actual feedback cancellation, thus preventing artifact propagation to the final output.
3Productivity
If existing ML-based AFC systems are trained to directly predict the feedback-free signal, then computational training effort increases, but the systems fail to provide satisfactory trade-off between convergence rate and steady-state behavior
Solution Approach 1:
The patent segments the feedback cancellation task into two distinct stages: (1) training the ML model to estimate the feedback path impulse response from feedback signals, and (2) using the estimated impulse response to generate cancellation signals. This segmentation reduces training computational complexity while improving the convergence-steady state trade-off by focusing training on the more tractable impulse response estimation subtask.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method, performed by an electronic device, for training a machine learning (ML) model for use in a feedback control system of a hearing aid is provided. The method comprises executing a plurality of training iterations. Each training iteration of the plurality of training iterations comprises obtaining training data. The training data comprises a training input signal and a training processed signal. The training input signal comprises an external input signal component and a feedback input signal component. The external input signal component is indicative of sound from a known, simulated acoustic environment of the hearing aid. The feedback input signal component is indicative of acoustic and/or mechanical feedback originating from a feedback path of the hearing aid. The training processed signal is indicative of one or more applied processing algorithms to a training feedback corrected input signal. The training feedback corrected input signal is indicative of a feedback corrected version of the training input signal. Each training iteration of the plurality of training iterations comprises obtaining target data comprising a training feedback path transfer function representative of an impulse response of the feedback path of the hearing aid. Each training iteration of the plurality of training iterations comprises determining, based on the training data, an estimate of the training feedback path transfer function. Each training iteration of the plurality of training iterations comprises updating the ML model based on the target data and the estimate of the training feedback path transfer function. The ML model comprises, in the following order, a convolutional layer, a first fully connected (FC) layer, and a long-short term memory (LSTM) layer.