Open-canal (OC) hearing aids have gained remarkable popularity in the past couple of years. In fact, a recent survey found that roughly a quarter of the BTEs dispensed in 2005 were OC fittings.1 Several factors, including cosmetic appeal, reduced occlusion effect, and improved digital signal processing (DSP) technologies, have contributed to the rise of OC hearing systems.
One DSP technology that has particular significance in the context of OC hearing aids is feedback reduction. Sophisticated feedback-reduction algorithms are an integral part of OC hearing aids, since they allow an additional 8 to 15 dB of gain before entering the audible oscillatory state. This article will examine the nature of feedback in hearing aids, how feedback-reduction algorithms work, and how to measure their effectiveness.
THE GENESIS OF FEEDBACK
First, let's briefly review how feedback occurs in hearing aids. As shown in Figure 1, a simplified schematic representation of feedback in hearing aids, a portion of the acoustic signal emanating from the receiver leaks back to the microphone through the feedback path. Note that the presence of a feedback path alone is not sufficient for the hearing aid to squeal; a condition called Nyquist Stability Criterion has to be violated. Put simply, for sustained feedback to occur, the gain through the forward or direct path of the hearing aid must be greater than the attenuation through the feedback path (for a more thorough explanation, see Agnew, Kuk et al., and Kates2–4).
The tonal nature of the feedback signal implies that this condition is met at only at a few frequencies. The exact value of these frequencies depends upon: (1) the gain characteristics of the hearing aid, and (2) the attenuation characteristics of the feedback path, which depend on the individual's ear canal characteristics, the nature of the fit, and the surrounding environment (nearby objects, room reverberation, etc.).
In addition, the feedback path is dynamic, implying that a change in the environment around the hearing aid (e.g., bringing a telephone closer, wearing a hat, giving a hug) will alter the attenuation characteristics of the feedback path and correspondingly affect the tonal composition of the feedback signal. In other words, the frequencies where feedback occurs when a hearing aid wearer brings a telephone closer to his/her ear may not be the same as when the person puts on a hat or hugs someone.
As an example, you probably have checked to see if a hearing aid is working by cupping your hands around it to incite feedback. If you change the way you cover the hearing aid by repositioning your hands, you will notice a difference in the tonal quality of the feedback signal. Essentially, you have changed the characteristics of the feedback path, which results in a different set of frequencies contributing to the feedback.
So far we have talked only about the audible feedback condition. However, the stage prior to this audible oscillatory stage is also important from a sound quality point of view. In this sub-oscillatory stage, the feedback components are not strong enough to induce sustained oscillation, but they can impact the frequency response of the hearing aid, as shown in Figure 2. The frequency response of a hearing aid is plotted here in no-feedback and sub-oscillatory feedback conditions as blue and pink curves, respectively. Note the appearance of the peak around 1000 Hz and the increased jaggedness in the frequency response for the sub-oscillatory condition, resulting in a poorer sound quality.
Audible feedback occurs only at select frequencies.
The exact frequencies depend on the parameters related to fitting (amount of venting, ear canal characteristics, hearing aid gain curve, etc.) and the surrounding environment.
Changes in the environment can result in a change in the composition of the feedback signal.
Feedback can affect the hearing aid frequency response and hence its sound quality even when it is inaudible.
A simple solution to the feedback problem is to reduce the overall gain of the hearing aid. However, that is impractical as we want to ensure appropriate gain and audibility for the hearing-impaired listener. Advances in DSP hardware and algorithms have resulted in sophisticated feedback-reduction techniques that often obviate the need for gain reduction while reducing the feedback signal. There are several different flavors of feedback-reduction algorithms, but let's look at two of the most popular: notch filtering and phase-cancellation or phase-inversion algorithms.
As its name suggests, a notch filter attempts to notch out a single frequency from the input signal. Figure 3 shows the frequency response of a digital notch filter tuned to 2500 Hz. A hearing aid using notch filtering to reduce feedback typically uses feedback detectors that monitor the input signal for the presence of sustained oscillations. Once sustained oscillation is detected, a notch filter is designed on the fly to reduce the gain around that frequency region.
If feedback occurs at more than one frequency, multiple notch filters are employed. If the feedback signal changes in its frequency composition, the notch filters will adapt to this change by correspondingly changing the location of their dip frequency, hence the term roving notch filters.
The main limitation with this approach is that reducing gain around the feedback frequency also affects the input signal, and this may impact audibility around that frequency region. This is especially significant if the bandwidth of notch filter is wide and is further compounded when multiple notch filters are used.
Another approach to feedback reduction uses phase cancellation. But before discussing feedback-reduction algorithms based on phase cancellation, let's talk about another example you may be familiar with. You've probably heard of the Active Noise Cancellation (ANC) headphones from Bose or other companies. They reduce the background noise when someone is listening through the headphones in an aircraft or some other noisy environment. The ANC headphones contain a microphone that samples the noisy environment. This noise is inverted (i.e., shifted in phase by 180°), added to the desired signal, and delivered through the speakers. The idea is to have the background noise and its phase-shifted version cancel each other out, thus creating a quieter environment for listening to music and other signals through the headphones.
The premise behind feedback-cancellation algorithms is similar: to create a copy of the feedback component and add it out of phase to the input signal. Figure 4 illustrates this concept in hearing aids.
As shown in Figure 3, an internal feedback path is created within the DSP system of the hearing aid that attempts to mimic the attenuation and phase characteristics of the external feedback path. The internal feedback path is essentially a digital filter whose frequency and phase response match those of the external feedback path. If everything works perfectly, the output of the internal feedback path is a replica of the feedback signal. When it is subtracted from the composite microphone signal, it results in a feedback-free output signal.
Unlike the ANC headphones described above, feedback phase-cancellation algorithms do not have the luxury of recording the feedback signal separately and using it to reduce the feedback in the input signal. Thus a major challenge in engineering a feedback-cancellation algorithm is the estimation of the internal feedback path such that the algorithm does not remove any components from the input signal. So, how does the DSP system in the hearing aid create this internal feedback path? In two ways:
- The hearing aid may explicitly measure the feedback path characteristics by playing back a broadband or swept sine stimulus through the receiver and measuring the response at the microphone. Several different hearing aids perform this open-loop gain measurement as part of detecting the presence of feedback and adjusting the gain margins accordingly. For example, Figure 5 shows a screenshot from Oticon's Genie fitting software. Pressing the “Start” button enables the fitting software to present a series of tones through the hearing aid receiver and measure the resulting response at the microphone. If feedback is detected, gain around that particular frequency region is reduced.
- A few hearing aids (e.g., ReSound Metrix, Starkey Axent) use the information obtained through open-loop gain measurements to initialize the digital filter that represents the internal feedback path.
- The second way is to dynamically adjust the coefficients of a digital filter that represents the internal feedback path such that the output of the subtraction block is devoid of the feedback components. Before you say “Huh?,” let me explain.
The key to understanding this approach is temporal correlation properties of the feedback signal and the desired input signal. Figure 6 shows the correlation plots of a speech signal (A boy fell from the window from the HINT database) and a sinusoidal signal. The correlation value on the Y-axis of these plots represents the degree of self-similarity in the input signal with respect to time. We can see from these two plots that the correlation value after a time delay will be much higher for a periodic signal such as a pure tone than a speech signal with time-varying spectral content.
This principle is exploited in estimating the internal feedback path. The response of the digital filter that represents the internal feedback path is adjusted such that the correlation between the microphone input and the amplifier output is minimized. Since the correlation is dominated by tonal components in the presence of feedback, such a minimization process will result in the reduction of the feedback components.
However, this reduction of correlation also means that any pure-tone-like components in the input signal may also be reduced by the feedback algorithm!
PERFORMANCE OF FEEDBACK-REDUCTION ALGORITHMS
How well do the feedback-reduction algorithms work? Figure 7 shows the experimental setup that our lab designed to measure their performance.
We use the Brüel & Kjær Head and Torso Simulator (Hobbes) with a telephone handset positioner. The telephone handset positioner allows several degrees of freedom in positioning the telephone handset. In addition, the position of the telephone can be precisely noted, which helps in taking multiple measurements with different hearing aids.
Figure 7 shows that the generic cellular phone handset was placed and positioned very close to the hearing aid. The hearing aid was programmed to fit a 50-dB flat loss and was verified using the Audioscan Verifit. Sentences from the HINT database were played back at 65 dB SPL from a loudspeaker positioned one meter in front of the manikin.
Figure 8 provides an example of the feedback-cancellation algorithm in action. In both panels, time is represented on the X-axis and frequency on the Y-axis, and the color represents the level at a given time/frequency. The top panel in Figure 8 shows the spectrogram of the hearing aid output with the feedback cancellation turned off. In this panel, a horizontal line is clearly visible representing the feedback frequency at 5600 Hz. The bottom panel shows the spectrogram of the hearing aid output with the feedback canceler turned on, where it is evident that the horizontal line representing the dominant feedback component has been reduced significantly. Also note in these pictures that the spectrum levels of the speech signal (as indicated by the red and yellow regions) have not changed appreciably with the feedback canceler turned on. This implies that the hearing aid is not reducing its gain to counteract the feedback signal.
Similar results can be obtained from other hearing aids, which indicates that feedback-reduction strategies based on phase cancellation do work. But, are all feedback cancelers the same? By no means. Different hearing aid manufacturers have their own spin on feedback-cancellation strategy. These algorithms are proprietary, so, while a general description of their strategy is available on manufacturers' web sites, details are not readily available. Perhaps a more pertinent question is: Have there been any studies comparing the effectiveness of feedback-cancellation algorithms? If we glean the peer-reviewed publications in audiology-related journals, we see a very few comparative studies:
Freed and Soli investigated the comparative performance of feedback cancelers in commercial hearing aids under various feedback conditions.5 In this purely electroacoustic study, feedback-cancellation performance was measured in terms of the added stable gain and the algorithm's ability to suppress peaks in the frequency spectrum, quickly react to a change in the external feedback path, and preserve tonal components in the desired signal. The metrics did show a varied performance across different devices. However, there are two issues with the study: (1) There were some questions on the repeatability of the metrics, and (2) it is unclear if the electroacoustic measures would have any perceptual relevance.
Greenberg, Zurek, and Brantley compared the performance of three adaptive feedback-reduction algorithms.6 These algorithms were implemented in a wearable DSP system and their performance was compared in terms of the added stable gain. In addition, ratings of speech quality were obtained from hearing-impaired listeners with and without the feedback-cancellation algorithm. Results from their study demonstrated that one particular algorithm performed the best on both objective and subjective measures. It is unclear if this particular algorithm is implemented in a commercial hearing aid.
Clearly we need more research in this area. Our lab is working on evaluating the impact of feedback-cancellation algorithms on speech and music quality. We are collecting speech and music data from modern hearing aids under various controlled feedback conditions with the feedback canceler turned on/off. Our plan is to measure the performance of the feedback-reduction algorithms using both electroacoustic measurements and perceptual sound quality experiments with hearing-impaired listeners.
Feedback-cancellation technologies do work. However, there are a few things about them to keep in mind:
- They may remove tonal components from the input signal. This becomes significant when the hearing aid processes a harmonically rich signal such as music. As you can imagine, removing tonal components from the music signal will have an adverse effect on the sound quality. Some manufacturers deal with this by providing a “music mode” for the feedback-cancellation algorithm. There is some evidence that the “music mode” does a better job at preserving the harmonic relationships of the music signal.5 However, both the added stable gain and the reaction time to a change in the feedback path are negatively affected by the “music mode.”
- Feedback-cancellation technologies may not respond well to a change in the feedback path. We are noticing in our lab that some hearing aids do not respond properly to a change in the feedback path, while some others take a long time to readjust (see Figure 9).
- Figure 9 shows the spectrogram of the hearing aid output in response to a string of ten HINT sentences played out at 65 dB SPL. During the playback, the feedback path is suddenly changed after the first two HINT sentences were finished. This change resulted in strong feedback components, as shown by the dark red regions in the figure. As evident from this figure, these feedback components persisted for a significant period of time.
- Those hearing aids where open-loop gain measurements have been performed and the feedback path initialized may provide less added stable gain in conditions where the feedback path differs significantly from the initialized one.
Open-canal hearing aids owe a large portion of their success to the performance of the feedback-cancellation algorithms. These algorithms allow an additional 8 to 15 dB of gain, which is invaluable for open fits.
Key things to keep in mind with respect to feedback-cancellation technologies are: (1) performance varies from device to device, (2) performance can be poor in dynamic situations, and (3) feedback-cancellation algorithms may remove tone-like components from the desired input signal such as music.
Funding provided by the Oticon Foundation and the Natural Sciences and Engineering Research Council (NSERC) of Canada is gratefully acknowledged.