Hearing Aid Feedback Control Using ML Feedback Path Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
State-of-the-art acoustic feedback cancellation techniques in hearing aids face challenges in balancing rapid convergence and low steady-state error, leading to compromised sound quality and speech intelligibility, particularly in real-world conditions.
Innovation Solution
A method for training a machine learning 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, addressing the trade-off between convergence rate and steady-state behavior.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If state-of-the-art adaptive filtering algorithms are used for acoustic feedback cancellation, then the system can operate with existing technology, but the convergence rate is slow and steady-state error is high, compromising sound quality and speech intelligibility
Solution Approach 1:
The patent introduces an intermediate representation (feedback path impulse response) as a mediator between the raw audio signals and the feedback cancellation process. By estimating this intermediate parameter first through a trained neural network, the system can then apply feedback cancellation more effectively. This intermediary step allows the system to achieve both fast convergence and low steady-state error, as the neural network rapidly estimates the feedback path characteristics which are then used for precise cancellation.
Solution Approach 2:
The patent replaces traditional adaptive filtering algorithms (mechanical signal processing approach) with a neural network-based system (learning-based approach). The neural network is trained offline to learn the complex nonlinear relationships in acoustic feedback paths, and during operation, it rapidly estimates feedback parameters without requiring iterative adaptive filtering. This substitution enables significantly faster convergence while maintaining or improving cancellation accuracy.
2Productivity
If existing ML-based AFC systems directly predict the feedback-free signal, then the system can provide feedback cancellation, but artifacts are introduced in the feedback-free signal when prediction is not optimal
Solution Approach 1:
Instead of directly predicting the feedback-free signal, the system uses an intermediate representation (feedback path impulse response) as a mediator. The neural network estimates this intermediate parameter, which then serves as the basis for feedback cancellation. This indirect approach avoids the artifacts introduced by direct prediction, as the intermediate representation provides a more stable and accurate foundation for cancellation, improving both signal quality and reliability.
3Measurement precision
If existing ML-based AFC systems are trained to directly predict the feedback-free signal, then the system can provide feedback cancellation, but the training is computationally intensive
Solution Approach 1:
The patent extracts and separates the training process from the operational process. The neural network is trained offline using simulated acoustic environments to learn feedback path characteristics. During actual operation, the pre-trained network only needs to estimate the feedback path impulse response, which is computationally much lighter than full feedback cancellation training. This extraction of the intensive training phase reduces real-time computational requirements while maintaining high cancellation accuracy.
Solution Approach 2:
The system performs preliminary training action offline using simulated data before actual deployment. By pre-training the neural network on diverse acoustic environments in advance, the system accumulates knowledge about feedback path variations without incurring computational costs during real-time operation. This preliminary action enables the system to handle real-world variations efficiently during operation with minimal computational burden.
Data Source
AI summary
Disclosed herein are embodiments of methods, performed by an electronic device, for training a machine learning (ML) model for use in a feedback control system of a hearing aid. The method can include executing a plurality of training iterations using training data including a training input signal and a training processed signal. Each training iteration of the plurality of training iterations can include updating the ML model based on a target data and an estimate of the training feedback path transfer function. Embodiments of the ML model includes, in the following order, a convolutional layer, a first fully connected (FC) layer, and a long-short term memory (LSTM) layer.