Keywords

1 Introduction

Age related hearing loss (presbycusis) makes it difficult to understand speech in noisy environments and multi-speaker environments (Moore 1995; Humes and Dubno 2010). There are several factors involved in presbycusis: loss of frequency selectivity, recruitment, and loss of temporal fine structure (Moore 2007). The isolation and measurement of these different aspects of hearing impairment (HI) in elderly listeners is not straightforward, partly because they often have multiple auditory problems, and partly because they may also have more central cognitive problems including depression. Together the problems make it difficult to obtain sufficient behavioral data to isolate different aspects of HI (Jerger et al. 1989; Lopez et al. 2003). We have developed a computational model of HI and an HI simulator that allow us to perform behavioural experiments on cochlear compression—experiments in which young normal hearing (NH) listeners act as “patients” with isolated, sensory-neural hearing losses of varying degree.

The HI simulator is essentially a high-fidelity synthesizer based on the dynamic, compressive gammachirp filter bank (dcGC-FB) (Irino and Patterson 2006). It analyses naturally recorded speech sounds with a model of the auditory periphery that includes fast-acting compression and then resynthesizes the sounds in a form that counteracts the compression in a NH listener. They hear speech with little distortion or background noise (Irino et al. 2013) but without their normal compression. This makes it possible to investigate recruitment phenomenon that are closely related to loss of cochlear compression (Bacon et al. 2004) using NH listeners who have no other auditory problems. Specifically, we describe an experiment designed to examine how the loss of cochlear compression affects the recognition of relatively soft syllables occurring in the presence of louder, flanking syllables of varying levels. The aim is to reveal the role of cochlear compression in multi-speaker environments or those with disruptive background noise.

2 Method

Listeners were presented three-syllable “nonsense words” and required to identify the second syllable. Between trails, the level of the second syllable within the word was varied (“dip” condition) or the level of the entire word was varied (“constant” condition). The difference between second-syllable recognition in the two conditions reveals the effect of compression. The HI simulator is used to adjust syllable level to simulate the hearing of 80 year old listeners with “normal” presbycusis (80 year 100 %) or a complete loss of compression (80 year 0 %).

2.1 Participants

Ten listeners participated in the experiment (four males; average age: 23.6 years; SD: 4.8 years). None of the listeners reported any history of hearing impairment. The experiment was approved by the ethics committee of Wakayama University and all listeners provided informed consent. Participants were paid for their participation, except for the first and third authors who also participated in the experiment.

2.2 Stimuli and Procedure

The stimuli were generated in two stages: First, three-syllable nonsense words were composed from the recordings of a male speaker, identified as “MIS” (FW03) in a speech sound database (Amano et al. 2006), and adjusted in level to produce the dip and constant conditions of the experiment. Second, the HI simulator was used to modify the stimuli to simulate the hearing of two 80 year old listeners, one with average hearing for an 80 year old and one with a complete loss of compression. There was also a normal-hearing control condition in which the level was reduced a fixed 30 dB.

In the first stage, nonsense words were composed by choosing the second syllable at random from the 50 Japanese syllables presented without parentheses in Table 1. This is the list of syllables (57-S) recommended by the Japanese Audiological Society (2003) for use in clinical studies to facilitate comparison of behavioural data across studies. The first and third syllables in the words were then selected at random from the full set of 62 syllables in Table 1. The set was enlarged to reduce the chance of listeners recognizing the restriction on the set used for the target syllable. Syllable duration ranged from 211 to 382 ms. There was a 50-ms silence between adjacent syllables within a word.

Table 1 The 62 Japanese syllables used in the experiment using the international phonetic alphabet. The syllables in parentheses are also Japanese syllables but not included in list 57-S. All 62 were used in the draw for the first and third syllables. The draw for the second syllable was restricted to the 50 syllables not enclosed in parentheses
Full size table

The nonsense words were then adjusted to produce the two level conditions illustrated in Fig. 1: In the “dip” condition (left column), the level of the second syllable was 70, 60, 50 or 40 dB, while the first and third syllables were fixed at 70 dB. In the “constant” condition (right column), all three syllables were set to 70, 60, 50 or 40 dB.

Fig. 1
figure 1

Schematic of the experimental paradigm. The levels of the nonsense words were adjusted to produce stimuli for a the Dip condition and b the Constant condition. The 70-dB stimuli were the same in the two conditions

Full size image

In the second stage, the stimuli were processed to simulate:

  1. 1.

    the average hearing level of an 80 year old person (audiogram ISO7029, ISO/TC43 2000) simulated with 100 % compression (80 year 100 %),

  2. 2.

    the average hearing level of an 80-year-old simulated with 0 % compression (80 year 0 %),

  3. 3.

    normal adult hearing with a 30-dB reduction across the entire frequency range (NH ‑30 dB).

To realize the 80 year 100 % condition, the output level of each channel was decreased until the audiogram of the dcGC-FB (normal hearing) matched the average audiogram of an 80 year old. To realize the 80 year 0 % condition, an audiogram was derived with the dcGC-FB having no compression whatsoever, and then the output level of each channel was decreased until the audiogram matched the audiogram of an average 80 year old. In the compression-cancellation process, loss of hearing level was limited so that the audiogram did not fall below the hearing level of an average 80 year old. The compression was applied using the signal processing system described in Nagae et al. (2014).

Each participant performed all conditions—a total of 1050 trials of nonsense words: 50 words × 7 dip/constant conditions × 3 hearing impairment simulations. The participants were required to identify the second syllable using a GUI on a computer screen. The stimuli were manipulated, and the results collected using MATLAB. The sounds were presented diotically via a DA converter (Fostex, HP-A8) over headphones (Sennheiser, HD-580). The experiment was carried out in a sound-attenuated room with a background level of about 26 dB LAEq.

3 Results

For each participant, the four percent-correct values for each level condition (dip or constant) were fitted with a cumulative Gaussian psychometric function to provide a Speech Reception Threshold (SRT) value for that condition, separately in each of the three simulated hearing-impairment conditions. The cumulative Gaussian was fitted by the bootstrap method (Wichmann and Hill 2011a; Wichmann and Hill 2011b) and SRT was taken to be the sound pressure level associated with 50 % correct recognition of the second syllable. The SRT in the constant condition with the NH ‑30 dB hearing impairment was taken as standard performance on the task. The difference between this standard SRT and the SRTs of the other conditions are the experimental results (ΔSRTs) for each participant. Figure 2a shows the mean and standard deviation over participants of ΔSRT for each condition.

Fig. 2
figure 2

a The difference in SRT (ΔSRT) between the constant condition of the NH ‑30 dB simulation and the SRTs of the other conditions. b The SRT difference between the dip and constant conditions for each hearing impairment condition. Asterisks (*) show significant differences (α = 0.05) in a multiple comparison, Tukey-Kramer HSD test. Error bars show the standard deviation

Full size image

A two-way ANOVA was performed for the dip/constant conditions and the hearing impairment conditions combined, and it showed that the manipulations had a significant effect on recognition performance (ΔSRT) (F(5, 54) = 23.82, p < 0.0001). Significant, separate main effects were also observed for the dip/constant condition and the hearing impairment condition (F(1, 54) = 41.99, p < 0.0001; F(2, 54) = 36.11, p < 0.0001, respectively). The interaction between the two factors was not significant (α= 0.05). Multiple comparisons with the Tukey-Kramer HSD test were performed. There were significant differences between all of the hearing impairment conditions (α= 0.05).

ΔSRT provides a measure of the difficulty of the task, and the statistical analysis shows that the different hearing impairment simulations vary in their difficulty. In addition, ΔSRT for the dip conditions is larger than ΔSRT for constant conditions for all of the hearing impairment simulations. This difference in difficulty could be the result of forward masking since the gap between offset of the first syllable and the onset of the target syllable is less than 100 ms, and the third syllable limits any post processing of the second syllable. The dip condition would be expected to produce more masking than the constant condition, especially when the target syllable is sandwiched between syllables whose levels are 20 or 30 dB greater. The constant condition only reveals differences between the different hearing impairment simulations. This does, however, include differences in audibility.

Figure 2b shows the difference between SRT in the dip and constant conditions; that is, ΔSRT relative to the constant condition. It provides a measure of the amount of any forward masking. A one-way ANOVA was performed to confirm the effect of hearing impairment simulation (F(2, 29) = 6.14, p = 0.0063). A multiple comparisons test (Tukey-Kramer HSD) showed that ΔSRT in the 80 year 0 % condition was significantly greater than in the 80 year 100 % condition, or the NH ‑30 dB condition (α= 0.05).

4 Discussion

It is assumed that forward masking occurs during auditory neural processing, and that it operates on peripheral auditory output (i.e. neural firing in the auditory nerve). Since the compression component of the peripheral processing is normal in both the 80 year 100 % condition and the NH ‑30 dB condition, the level difference between syllable 2 and syllables 1 and 3 should be comparable in the peripheral output of any specific condition. In this case, ΔSRT provides a measure of the more central masking that occurs during neural processing. When the compression is cancelled, the level difference in the peripheral output for syllable 2 is greater in the dip conditions. If the central masking is constant, independent of the auditory representation, the reduction of the output level for syllable 2 should be reflected in the amount of masking, i.e. ΔSRT. This is like the description of the growth of masking by Moore (2012), where the growth of masking in a compressive system is observed to be slower than in a non-compressive system. Moore was describing experiments where both the signal and the masker were isolated sinusoids. The results of the current study show that the effect generalizes to speech perception where both the target and the interference are broadband, time-vary sounds.

There are, of course, other factors that might have affected recognition performance: The fact that the same speaker provided all of the experimental stimuli might have reduced the distinctiveness of the target syllable somewhat, especially since syllable 2 was presented at a lower stimulus level in the dip conditions. The fact that listeners had to type in their responses might also have decreased recognition performance a little compared to everyday listening. It is also the case that our randomly generated nonsense words might accidentally produce a syllable sequence that is similar to a real word with the result that the participant answers with reference to their mental dictionary and makes an error. But, all of these potential difficulties apply equally to all of the HI conditions, so it is unlikely that they would affect the results.

This suggests that the results truly reflect the role of the loss of compression when listening to speech in everyday environments for people with presbycusis or other forms of hearing impairment. It shows that this new, model-based approach to the simulation of hearing impairment can be used to isolate the compression component of auditory signal processing. It may also assist with isolation and analysis of other aspects of hearing impairment.

5 Summary

This paper reports an experiment performed to reveal the effects of a loss of compression on the recognition of syllables presented with fluctuating level. The procedure involves using a filterbank with dynamic compression to simulate hearing impairment in normal hearing listeners. The results indicate that, when the compression of normal hearing has been cancelled, it is difficult to hear syllables at low sound levels when they are sandwiched between syllables with higher levels. The hearing losses observed in patients will commonly involve multiple auditory problems and they will interact in complicated ways that make it difficult to distinguished the effects of any one component of auditory processing. The HI simulator shows how a single factor can be isolated using an auditory model, provided the model is sufficiently detailed.