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

In feedback cancellation in hearing aids, an adaptive filter is used to model the feedback path. The output of the adaptive filter is subtracted from the microphone signal to cancel the acoustic and mechanical feedback picked up by the microphone, thus allowing more gain in the hearing aid. In general, the feedback cancellation filter adapts on the hearing-aid input signal, and signal cancellation and coloration artifacts can occur for a narrowband input. The goal in feedback cancellation is to design an algorithm that is computationally efficient, provides 10 dB or more of additional amplifier headroom, and is free from processing artifacts.This chapter begins with a steady-state analysis of the effects of acoustic feedback on the hearing-aid response. This analysis illustrates the factors that can affect the accuracy of the adaptive filter and the filter convergence for systems adapting with or without a probe signal. The characteristics of the feedback path being modeled are then described. An additional concern in designing signal processing for a hearing aid are the processing constraints imposed by a low-power portable device, and these concerns are discussed next. Two forms of constrained adaptation are then derived, and simulation results are used to give a comparison of the constrained adaptation with the conventional unconstrained approach for a system adapting without a probe signal. Further improvements in feedback cancellation performance, particularly a reduction in audible processing artifacts, can be achieved by using the filtered-X algorithm, and this approach is described next.The performance of feedback cancellation is ultimately limited by the characteristics of the room in which the listener is located since the room reflections form part of the feedback path. The chapter concludes with a study of the feedback path in a room along with the feedback cancellation performance for an 8-, 16-, or 32-tap adaptive FIR filter in series with a five-pole nonadaptive filter, and the performance limitations imposed by the room reverberation are discussed.KeywordsImpulse ResponseFeedback PathAdaptive FeedbackAcoustic FeedbackRoom ReverberationThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

In this paper, we propose a novel solution for continuous acoustic feedback neutralization in the digital hearing aids. The proposed method is based on two adaptive filters working in tandem. The error signal of the first adaptive filter is used as a desired response for the second adaptive filter being excited by a low-level (constant) probe signal. Initially at the start up, the first adaptive filter gives a fast convergence, however, its input and desired response are correlated with each other and it may converge to a biased solution at the steady-state. On the other hand, the second adaptive filter, though converging slowly being excited by a low-level probe signal, would give a good steady-state estimate of the acoustic feedback path. An appropriate delay is inserted at the output of the hearing aid, and based-on this delay-based processing, a novel strategy is proposed to exchange the weights between the two adaptive filters. The proposed strategy is such that the misconvergence of the first adaptive filter is avoided, as well as the initial convergence of the second adaptive filter is improved. Extensive computer simulations are carried out which demonstrate the effectiveness of the proposed method.

Acoustic feedback in hearing aids occurs due to the coupling between the hearing aid loudspeaker and microphones. In order to reduce acoustic feedback, adaptive filters are often used to estimate the feedback path. To increase the convergence speed and decrease the computational complexity of the adaptive algorithms, it has been proposed to split the acoustic feedback path into a time-invariant fixed part and a time-varying variable part. A key question of this approach is how to determine the fixed part. In this paper, two approaches are investigated: (1) a digital filter design approach that makes use of the signals of at least two hearing aid microphones and (2) a defined physical location approach using an electro-acoustic model and the signals of one hearing aid microphone and an additional ear canal microphone. An experimental comparison using measured acoustic feedback paths showed that both approaches enable one to reduce the number of variable part coefficients. It is shown that individualization of the fixed part increases the performance. Furthermore, the two approaches offer solutions for different requirements on the effort to a specific hearing aid design on the one hand and the effort during the hearing aid fitting on the other hand.

Feedback is a problem in hearing aids which will cause signal degradation and reduce the maximum applicable gain. More specifically, the advantages of open fittings (e.g., minimizing the occlusion effect) are limited by acoustic feedback. Feedback cancelation algorithms are used to overcome these limitations. For the development of such algorithms, the acoustic feedback path of the hearing aid must be known. The acoustic feedback path is not only affected by the outer sound field but by the individual anatomy and physiology as well. In order to quantify these different influences, feedback path measurements were performed on 20 human subjects. The measurements included different static conditions as well as dynamic ones (i.e., repetitive movements were performed during the measurement). Since the sound pressure level must be limited in measurements on human subjects, a valid identification of the feedback path is difficult in many cases, due to a low signal to noise ratio. Therefore, all measurements were performed reciprocally in addition to the direct measurements. Results show that yaw movements only have a small influence on the acoustic feedback path, whereas changes of the outer sound field can have a substantial impact on the feedback path.

The computational complexity and convergence speed of adaptive feedback cancellation algorithms depend on the number of adaptive parameters used to model the acoustic feedback path. To reduce the number of adaptive parameters it has been proposed to decompose the acoustic feedback path as the convolution of a (time-invariant) common part and a (time-varying) variable part. Typically the problem of estimating all the required coefficients has been formulated as a least-squares optimization problem. In contrast, in this paper we propose to formulate the estimation problem as a minmax optimization problem and show how this is associated with the maximum stable gain of a hearing aid. Experimental results using measured acoustic feedback paths from a two-microphone behind-the-ear hearing aid show that the proposed minmax optimization outperforms the least-squares optimization in terms of maximum stable gain. Furthermore, the robustness of proposed common part decomposition for different feedback paths is evaluated.

While several proactive acoustic feedback (Larsen-effect) cancellation schemes have been presented for speech applications with short acoustic feedback paths as encountered in hearing aids, these schemes fail with the long impulse responses inherent to, for instance, public address systems. We derive a new prediction error method (PEM)-based scheme (referred to as PEM-AFROW) which identifies both the acoustic feedback path and the nonstationary speech source model. A cascade of a short- and a long-term predictor removes the coloring and periodicity in voiced speech segments, which account for the unwanted correlation between the loudspeaker signal and the speech source signal. The predictors calculate row operations which are applied to prewhiten the speech source signal, resulting in a least squares system that is solved recursively by means of normalized least mean square or recursive least squares algorithms. Simulations show that this approach is indeed superior to earlier approaches whenever long acoustic channels are dealt with

An improved block adaptive system for effective feedback cancellation in hearing aids

Acoustic feedback is a serious problem in hearing aids. The feedback can result in whistling sound and make people who have severe hearing loss feel terrible. Traditional LMS and normalized LMS algorithm have been studied before, but these algorithms do not provide satisfactory performance for reducing feedback oscillation in hearing aids. So a band-limited NLMS algorithm is proposed. In order to determine which adaptive filter used is more suitable, we use these two kinds scheme to simulation and compare the Signal-to- Feedback-Ratio (SFR) of these two schemes, the band-limited algorithm has better performance and the output speech's quality is better. The commonly used adaptive filter is FIR filter, but it has more coefficients and high computational complexity, so a feedback reduction subsystem based on IIR adaptive filters is proposed. Its coefficients are fewer and computational cost is lower. In this case, the IIR filter is adapted with a modified version of the NLMS algorithm for this kind of filters. It uses fewer coefficients and obtains better performance.

Adaptive feedback cancellation in hearing aids

In adaptive feedback cancellation the computational complexity and the convergence speed are determined by the number of adaptive parameters used to model the acoustic feedback path. Therefore it has been proposed to reduce the number of adaptive parameters by modeling the feedback path as the convolution of a time-invariant common part and a time-varying variable part. While previous approaches have modeled the common part either using only poles or using only zeros, in this paper we propose to use a common pole-zero model and present an iterative method to compute the common poles and zeros. Using measured acoustic feedback paths from a two-microphone behind-the-ear hearing aid it is shown that the proposed model enables either to increase the modeling accuracy given a fixed number of parameters of the variable part or to reduce the number of parameters of the variable part given a desired accuracy.

At present, the study of acoustic echo cancellation (AEC) is mainly based on the adaptive acoustic echo cancellation algorithm. It is proved that the echo signal of the near-end microphone must be compensated by the time delay to achieve the purpose of echo cancellation, and the accuracy of the delay estimation affecting the final effect of echo cancellation. This paper proposes a combination of the normalized minimum mean square algorithm (NLMS) and the time delay estimation model to solve the echo problem in hearing aids. In this paper, using generalized cross correlation (GCC) to estimate time delay. In addition, using the energy, Teager-Kaiser Energy Operator (TKEO) and the signal correlation coefficient as the threshold value to detect the howling signal in digital hearing aids. Finally, the simulation and experimental results are given. The experiment proves that the method has good effect.

The adaptive feedback canceller (AFC) based on the least mean square (LMS) or normalized LMS (NLMS) algorithm shows poor convergence for sparse impulse response, since they do not consider the characteristics of the impulse response. The acoustic feedback path of the hearing aid exhibits sparse characteristics consisting of few active coefficients and many non active (near to zero) coefficient values. This paper proposes an improved gradient controlled improved proportionate NLMS (IGC-IPNLMS) algorithm for cancelling the feedback phenomenon in hearing aids. The IGC-IPNLMS algorithm employs a variable convergence factor, an estimate of the gradient vector for allocating step size, the controlling parameter of the IGC-IPNLMS algorithm is also updated in each iteration based on the sparseness measure. Further, the reduction in computational complexity is achieved by using sparse partial update method along with the IGC-IPNLMS algorithm (SIGC-IPNLMS) for feedback cancellation. The simulation results show the superior performance of the proposed method for white noise and speech segment as input.

Publication in the conference proceedings of EUSIPCO, Antalya, Turkey, 2005

In adaptive feedback cancellation the convergence speed and the computational complexity depend on the number of adaptive param-eters used to model the acoustic feedback path. To improve the convergence speed and reduce the computational complexity, it has been proposed to model the acoustic feedback path as the convolution of a time-invariant common pole-zero part and a time-varying variable part. Previous approaches to estimate all the coefficients minimized the so-called equation-error which possibly suffers from poor estimation accuracy in the vicinity of prominent spectral regions, e.g., spectral peaks. In this paper we therefore propose to minimize the so-called output-error by using a Steiglitz-McBride-like iteration scheme. To ensure the stability of the estimated pole-zero filter a frequency domain constraint is used leading to a quadratic programming problem. Experimental results using measured impulse responses from a two-microphone behind-the-ear hearing aid show that the proposed estimation scheme outperforms the existing estimation scheme in terms of modeling accuracy.

In this contribution, we present the noise management features of the Open Speech Platform (OSP) for hearing aid (HA) research. OSP includes basic HA modules (i) subband decomposition, (ii) wide dynamic range compression and (iii) adaptive feedback cancellation; a baseline single-channel speech enhancement (SE) based on Wiener filtering for noise subtraction. We extended OSP noise management to include a generalized sidelobe cancellation (GSC) beamforming between left and right channels, followed by the SE. Time is the scarcest resource, followed by central processing unit (CPU) resources in commercial HAs. With the proposed GSC + SE approach, we were both time and CPU limited. Release 2019a of OSP has 5.6 ms end-to-end latency and GSC requires an additional 5 ms, putting us above the 10 ms requirement. The wearable device of OSP has 4 cores, C0 – C3. C0 is used for all non-realtime tasks such as kernel, embedded web server, etc. and remaining cores are used for realtime tasks. Naive realization of GSC results in one or more cores not meeting the realtime constraints, resulting in audible artifacts. We present optimizations to meet time and CPU budgets; and preliminary objective and subjective results of the proposed system.