A Real Time Hearing Loss Simulator

Summary Several hearing loss simulators ( HLS ) have been developed to demonstrate the e ﬀ ects of hearing loss on auditory perception to normal hearing (NH) listeners, and to facilitate prediction of the perception of sound products by hearing impaired customers. This paper describes a real-time HLS based on an inverse, compressive GammaChirp (GC) ﬁlterbank, and how it was used to temporarily handicap NH listeners participating in a traditional notched-noise (NN) masking experiment (e.g. [1]) with a 2-kHz signal frequency. Sets of NN thresholds were obtain with a wide range of symmetric and asymmetric notches at two noise spectrum levels while participants listened to the sounds presented both with and without the HLS . The NN data were used to derive auditory ﬁlter shapes and input/output (IO) functions, which demonstrate that the HLS can simulate the elevation of pure tone threshold and the ﬂattening of the input/output function commonly observed in sensory-neural hearing loss. the


Introduction
Simulation of sensory processing disorders provides a powerful tool for investigating hearing itself and hearing loss. In the past, there have been twom ain approaches: The first wase quivalent-threshold masking, designed to simulate the reduced performance of hearing impaired listeners in one task or another,without regard for the quality of the perceivedsound (see Lum and Braida for areview [2]). Forexample, to simulate specificaudiometric losses, theyoften simply mixed aloud broadband band noise with the shape of the audiogram to the signal producing atotally different experience from what the impaired listener heard. In the second approach, theym ade an attempt to simulate the actual perception of the hearing impaired listener. In the simplest case, theyr educed the levelo ft he sound with alinear FFT-filter having afrequencyresponse close to that of the target audiogram. This wasu sed to demonstrate the consequences of sensory hearing impairment to the general public. But this linear attenuation of the signal is ap oor imitation of what happens when someone loses the active process of the peripheral auditory system and its associated gain. More recently,some integrated models of hearing impairment have been proposed [3,4], which were intended to simulate all aspects of moderate, sensory-

Model of hearing impairment
The principle of the current hearing loss simulator is essentially the same as that of the simulator developed by Irino and colleagues, referred to as an inverse, dynamic compressive Gammachirp (dcGC)auditory filter [4,5,6]. It wasu sed by Matsui et al [7] to simulate the effect of hearing loss in as yllable perception task. However, they did not derive either the auditory filter or the input/output function of the cochlea using the simulator.A st he name suggests, the inverse dcGC simulator wasdesigned to cancel the natural compression of the normal hearing listener. In the current dcGC model, cochlear compression is simulated in three stages: 1) The signal is filtered into 32 bands using abank of passive GammaChirp (pGC)filters. 2) The levelatthe output of each pGC filter is estimated. 3) The levelisused to control the center frequencyofahigh-pass asymmetry function (HP-AF) that represents the active mechanism in that filter band. The center frequencyofthe HP-AF decreases as the output levelofthe pGC increases, reducing filter gain and increasing filter bandwidth in the process. Thus, as in the cochlea, the gain is maximal at lowlevels and minimal at high levels, and the system provides fast acting compression overalarge dynamic range, separately in each dcGC band.
To cancel the natural compression of the dcGC filterbank, the hearing loss simulator HLS applies asecond ver- Figure 1. HLS signal processing: 1) pGG filters; 2) channelby-channel levele stimation; 3) calculation of HP-AF leveldependent filter coefficients and application of gain in each band; 4) time reversed pGC to cancel the delay group of each band; 5) passive gain to add apassive hearing loss (not used here); 6) sum all bands to re-synthesize.
sion of the active mechanism in reverse (see figure 1),that is, the center frequencyofthe second HP-AF increases as the levelout of the pGC increases. In this way, the simulator acts as an inverse compressor in each frequencyband, in aw ay that should cancel the natural compression of a normal listener.
The processing is done with amix of python and Open Computing Language (opencl). All filter coefficients are designed with acascade of biquad filters. The HP-AF coefficients are computed in advance for all levels and the resulting gains stored in alookup table. All steps are computed sample by sample in each band at 44.1 kHz; the use of aGraphics Processing Unit (GPU)a nd opencl makeit possible to process the 64 bands associated with binaural audio streams in real time. This means the hardware version of the simulator can be inserted in anyaudio system to simulate asensory-neural hearing impairment. The HLS software can be downloaded as an open source project 1 .

Notched-Noise Experiment
Af orm of GC HLS has previously been used to simulate the performance of ag roup of hearing impaired (HI) listeners on as peech-in-noise task [8]. The average audiogram of the HI group wasused to fit the HLS for the normal hearing listeners. It showed the presence of amoderate hearing loss that, in turn, explained their speech intelligibility deficit. However, it wasnot clear whether the deficit wasentirely attributable to their hearing losses or whether it wasatleast partially due to amore general deterioration of the signal. To resolvethe ambiguity and validate the current GC HLS,w ed esigned aN Ne xperiment, centered at 2kHz, to measure the effect of the HLS on absolute threshold, auditory filter shape, and the IO function of ag roup of normal hearing (NH) individuals, making adirect comparison with and without the HLS.Adetailed description of the NN experiment and the derivation of auditory filter shape with aGCfilter model is presented in Patterson et al (2003) [9].

Methods
Six young, normal hearing listeners were tested in their best ear,h aving giveni nformed consent prior to the start of the experiment.
Twosets of NN thresholds were collected, one without the HLS (referred to as the ByPass condition)and one that included the HLS (referred to as the HLS condition). In the latter condition, the system wasset to simulate acomplete loss of compression in all bands. In this case, the HLS prediction for absolute threshold at 2kHz increases by about 37 dB SPL.
Absolute threshold at 2kHz wasmeasured using atwointerval, two-alternative,f orced-choice procedure with a2 -down, 1-up tracking paradigm. The intervals were 200 ms in duration, separated by 500 ms. The timing of the intervals wasindicated visually on acomputer display. One interval, randomly selected, contained a2 00 ms sinusoid. The task of the listener wast oi ndicate the intervalthat had this signal by ab utton press. The initial level of the tone was4 0dBS PL int he ByPass conditions and 77 dB SPLinthe HLS conditions. The initial step size was 8dB. It wasreduced to 4dBafter tworeversals, and to its final levelof2dB after 2more reversals. Threshold measurement wasterminated after 16 reversals. Threshold was taken to be the average of the last 12 reversals. The conditions were presented in random order.
The same experimental procedure wasused to estimate signal threshold in the twoN Nc onditions -w ith, and without, the HLS.T he only difference wast hat the 200ms notch noise waspresent in both intervals of each trial. The spectrum levelo ft he NN was2 5o r4 5dBS PL in the ByPass condition and 45 or 60 dB SPL in the HLS condition. The initial levelo ft he tone wass et to 30 dB above the spectrum levelo ft he NN in all conditions (i.e. 55 dB, 75 dB or 90 dB). The widths of the lower and upper noise bands were fixed at 400 Hz. The notch noise was generated by filtering awhite noise with a16th order butterworth bandpass filter to establish the extremities of the NN. The notch wasthen added using a16th-order Butterworth, band-reject filter.D epending on the condition, the notch wasp ositioned either symmetrically or asymmetrically about the signal frequency, 2kHz. Nine or ten symmetrical notches were used for each noise level. were grouped into three blocks (A,Band C) to provide for breaks in the testing; each block contained about the same number of notch widths and levels. Overall, 21 or 22 notch conditions were tested at each of 2noise levels with, and without, the HLS.
The stimuli were passed through the numerical optical output of an RME sound card. This output wasthen connected to the numerical optical input of the same RME sound card, and this wasthe input to the HLS.The output of the simulator waspresented monaurally to the best ear of the listener through aSennheiser HD250, linear II headphone. The stimuli were calibrated using aclass Asound levelm eter (Larson Davis 824)c onnected to an artificial ear (Larson Davis AEC101). The listeners sat in adoublewalled sound booth. The experimental paradigm wasformally approvedbyanational ethics committee (CPP Léon Bérard).

Results
The average threshold data for the ByPass and HLS parts of the experiment are plotted, as afunction of notch width, in the upper and lower panels of Figure 2, respectively. The upper and lower threshold curves in the HLS condition have very similar shapes to the upper and lower threshold curves in the ByPass condition, indicating that the effect of the HLS is basically what it should be -asophisticated, fast acting sound attenuator.Relative to overall level, widening the notch produces avery similar effect on threshold after the intervention of the HLS,and this is true for the asymmetric notches (green and magenta symbols)a sw ell as the symmetric notches (black symbols). The main difference between the twop atterns of threshold curves is that the range of thresholds obtained with the HLS is somewhat compressed relative to the pattern in the ByPass condition.
The average value of absolute threshold is shown by the black, horizontal dashed line in each panel; the value was1 0.0 dB SPL (std=3.50)i nt he ByPass condition and 47.0 dB SPL(std=1.24)inthe HLS condition. The difference, 37 dB, is exactly the change in absolute threshold predicted by setting the degree of compression to zero in the HLS.N ote, however, that whereas absolute threshold is about 5dBbelowthe lowest NN threshold in the ByPass condition, it is al ittle above the lowest NN threshold in the HLS condition. We return to the differences between the ByPass and HLS threshold values in the Discussion.
In order to derive the auditory filters, the notch-noise data have been fitted using the same P0 power spectrum model of masking as described previously in this issue [10] and earlier [11,12]. In each condition, the minimization of equation 4 [10]: provides the full set of parameters of the dcGC model which best predicts the data. Using these parameters, figure 3, presents 5a uditory filters, at 5i nput levels (every 10 dB in the range of the data), derivedw ith the dcGC model in the ByPass and HLS parts of the experiment in the upper parts of panels Aa nd B, respectively.T he blue lines showthat the auditory filter provides gain in the passband region in both the ByPass and HLS conditions. The errors between the threshold values predicted by this dcGC filter model showthat the model provides an accurate description of the ByPass data with an error equal to 2.32 dB as indicated in Figure 3, and also an reasonable description of the HLS data with an rms error equal to 3.17 dB. In this condition, the prediction of the minimum threshold is about 5dBbelowabsolute threshold, and the model predictions are alittle above the corresponding data at the wider notch widths. This discrepancyi sr eflected in the higher rms error in this condition than in the ByPass condition.
Note that the design maximizes the number of different NN conditions in the experiment, in preference to replicating asmaller number of conditions, and so the error in the GC fitsincludes the intra-individual variability (i.e. the error in individual threshold estimation).
The input/output (IO) functions and the bandwidth (BW) functions provided by the dcGC model are plotted, as afunction of stimulus level, belowthe corresponding filter shape plots in Figure 3. The blue portions of the IO and BW curves showestimates from roughly the same range of levels as the measured thresholds; the cyan sections showextrapolations to lower and higher levels.

Discussion
The IO function for the ByPass condition is strongly compressive,a se xpected, with as lope of 0.2 dB/dB for input levels around 60 dB SPL. The IO function for the HLS condition is much less compressive;the minimum is 0.41 dB/dB. The form is consistent with the loss of compression that would be expected from aHIlistener with a 37-dB hearing loss.
The BW values for filters in the levelrange of the threshold data (the blue portion of the BW function)are 1.6-2.0 times the normal ERB (ERB N )v alue [13] in the ByPass condition, and 1.3-2.2 times the ERB N value in the HLS condition. Part of the difference arises from the fact that the ERB N BW values were derivedwith aroexfilter-shape which has been shown to underestimate the actual width of the tip of the auditory filter (see [13], subsection IV.B).
It remains the case, however, that the average BW value in the HLS condition is somewhat smaller than that in the ByPass condition. This is because the HLS simulates the loss of gain in the HI by reducing the levelofthe stimuli (signal + maskers)p resented to the NH listeners in HLS condition. That is, the sounds are actually being presented to these NH listeners at amuch lower levelthan the CP input axis would suggest. In the ByPass condition, the stimuli are being presented at the stated CP input level. The BW of the NH listeners is greater at higher levels, so the BW values are greater in the ByPass condition than in the HLS condition. This does, however, mean that the HLS is limited to simulating the loss of gain in the HI; it does not simulate the increase in BW associated with the need to present stimuli at higher levels for the HI.

Conclusions
The HLS wasobserved to raise absolute threshold substantially and reduce compression, making the auditory system appear more linear.T hese changes are qualitatively consistent with the presence of afl at hearing loss of around 40 dB. Asimilar simulator [8] has been shown to produce areduction of intelligibility for speech presented in noise, similar to that observed with HI listeners. The results of the current experiment allowu st oc onclude, more generally, that the HLS illustrates the joint effects of reduced audibility and reduced compression commonly encountered in HI listeners.