Keywords

1 Introduction

Speech perception is a major communication challenge for people with hearing loss and with cochlear implants (CIs), particularly when the speech is embedded in background noise (e.g., Humes et al. 2002; Zeng 2004). Recent work has suggested that it is not so much the overall noise energy that limits speech perception in noise, as suggested by earlier work (French and Steinberg 1947; Kryter 1962; George et al. 2008), but rather the energy in the inherent temporal-envelope modulations in noise (Dubbelboer and Houtgast 2008; Jorgensen and Dau 2011; Stone et al. 2011, 2012; Jorgensen et al. 2013; Stone and Moore 2014). In a recent study (Oxenham and Kreft 2014) we examined the effects of inherent noise fluctuations in CI users. In contrast to the results from normal-hearing (NH) listeners, we found that CI users exhibited no benefit of maskers without inherent fluctuations. Further experiments suggested that the effective inherent noise envelope fluctuations were reduced in the CI users, due to the effects of current spread, or interactions between adjacent electrodes, leading to smoother temporal envelopes.

One remaining question is whether HI listeners exhibit the same loss of sensitivity to inherent noise fluctuations as CI users. If so, the finding may go some way to explaining why both HI and CI populations exhibit less masking release than NH listeners when additional slow fluctuations are imposed on noise maskers (e.g., Festen and Plomp 1990; Nelson and Jin 2004; Stickney et al. 2004; Gregan et al. 2013). On one hand, cochlear hearing loss is generally accompanied by a loss of frequency selectivity, due to loss of function of the outer hair cells; on the other hand this “smearing” of the spectrum occurs before extraction of the temporal envelope, rather than afterwards, as is the case with CI processing. Because the smearing is due to wider filters, rather than envelope summation, the resultant envelopes from broader filters still have Rayleigh-distributed envelopes with the same overall relative modulation power as the envelopes from narrower filters (i.e., the same area under the modulation power spectrum, normalized to DC; see Fig. 1). In contrast, with CIs, the envelopes derived from summing the envelope currents from adjacent electrodes are no longer Rayleigh distributed and have lower modulation power, relative to the overall (DC) power in the envelope (Hu and Beaulieu 2005).

Fig. 1
figure 1

Schematic diagram of the effect of broadening the filter from a bandwidth of W to a bandwidth of 2 W on the modulation spectrum of filtered Gaussian noise. The relative modulation power (area under the line) remains constant, but the area under the lines within the speech-relevant range (shaded rectangle) is reduced

Full size image

One factor suggesting that HI listeners may also experience less influence of inherent noise fluctuations is that the modulation spectrum is altered by broadening the filters: for an ideal rectangular filter, the modulation power of Gaussian noise after filtering has a triangular distribution, reaching a minimum of no power at a frequency equal to the bandwidth of the filter (Lawson and Uhlenbeck 1950). Although widening the filter does not alter the area under the modulation spectrum, it results in relatively less power at lower modulation frequencies (see Fig. 1). Given that low modulation frequencies are most important for speech, the relative reduction in modulation power at low modulation frequencies may reduce the influence of the inherent fluctuations for listeners with broader filters, due to hearing loss.

The aim of this experiment was to test the resulting prediction that hearing loss leads to less effect of inherent noise fluctuations on speech masking. We compared the results of listeners with cochlear hearing loss with the performance of young NH listeners and age-matched NH listeners. Performance was compared for roughly equal sensation levels (SL), and for equal sound pressure levels (SPL) to test for the effect of overall level on performance in NH listeners.

2 Methods

2.1 Listeners

Nine listeners with mild-to-moderate sensorineural hearing loss (4 male and 5 female; mean age 61.2 years) took part in this experiment. Their four-frequency pure-tone average thresholds (4F-PTA from 500, 1000, 2000, and 4000 Hz) ranged from about 25 to 65 dB HL (mean ~ 40 dB HL). Nine listeners with clinically normal hearing (defined as 20 dB HL or less at octave frequencies between 250 and 4000 Hz; mean 4F-PTA 7.6 dB HL), who were matched for age (mean age 62.2 years) and gender with the HI listeners, were run as the primary comparison group. In addition, a group of four young (mean age 20.5 years; mean 4F-PTA 2.8 dB HL) NH listeners were tested. All experimental protocols were approved by the Institutional Review Board of the University of Minnesota, and all listeners provided informed written consent prior to participation.

2.2 Stimuli

Listeners were presented with sentences taken from the AZBio speech corpus (Spahr et al. 2012). The sentences were presented to the HI listeners at an overall rms level of 85 dB SPL. The sentences were presented to the NH listeners at two different levels: in the equal-SL condition, the sentences were presented at 40 dB above the detection threshold of the speech for the NH listeners, similar to the sensation level (SL) of the 85-dB SPL speech for the HI listeners; in the equal-SPL condition, the speech was presented at 85 dB SPL. The sentences were presented in three different types of masker, as in Oxenham and Kreft (2014); see Fig. 2. The first was a Gaussian noise, spectrally shaped to match the long-term spectrum of the target speech. The second was comprised of 16 pure tones, approximately evenly spaced on a logarithmic frequency scale from 333 to 6665 Hz, corresponding to the center frequencies on a standard CI map for Advanced Bionics. The amplitudes of the tones were selected to produce the same long-term output of a 16-channel vocoder as the target speech (or Gaussian noise masker), assuming the same center frequencies. The third was comprised of the same 16 tones, but each tone was modulated independently with the temporal envelope of a noise masker, bandpass filtered using a vocoder filter with the same center frequency as the tone carrier. These three maskers produced equal amounts of speech masking in the CI users tested by Oxenham and Kreft (2014). The masker was gated on 1 s before the beginning of each sentence, and was gated off 1 s after the end of each sentence. The masker in each trial was a sample of a longer 25-s sound file, cut randomly from within that longer waveform. The speech and masker were mixed before presentation and the signal-to-masker ratios were selected in advance, based on pilot data, to span a range of performance between 0 and 100 % word recognition.

Fig. 2
figure 2

Representation of the three masker types used in the experiment. The three panels provide spectral representations of the noise (top), tone (middle), and modulated tone (bottom) maskers

Full size image

The speech and the masker were mixed and low-pass filtered at 4000-Hz, and were either presented unprocessed or were passed through a tone-excited envelope vocoder that simulates certain aspects of CI processing (Dorman et al. 1998; Whitmal et al. 2007). The stimulus was divided into 16 frequency subbands, with the same center frequencies as the 16 tone maskers. The temporal envelope from each subband was extracted using a Hilbert transform, and then the resulting envelope was lowpass filtered with a 4th-order Butterworth filter and a cutoff frequency of 50 Hz. This cutoff frequency was chosen to reduce possible voicing periodicity cues, and to reduce the possibility that the vocoding produced spectrally resolved components via the amplitude modulation. Each temporal envelope was then used to modulate a pure tone at the center frequency of the respective subband.

2.3 Procedure

The stimuli were generated digitally, converted via a 24-bit digital-to-analog converter, and presented via headphones. The stimuli were presented to one ear (the better ear in the HI listeners), and the speech-shaped noise was presented in the opposite ear at a level 30 dB below the level of the speech. The listeners were seated individually in a double-walled sound-attenuating booth, and responded to sentences by typing what they heard via a computer keyboard. Sentences were scored for words correct as a proportion of the total number of keywords presented. One sentence list (of 20 sentences) was completed for each masker type and masker level. Presentation was blocked by condition (natural and vocoded, and speech level), and the order of presentation was counterbalanced across listeners. The test order of signal-to-masker ratios was random within each block.

3 Results

The results from the conditions where the speech was presented at roughly equal sensation level to all participants (85 dB SPL for the HI group, and 40 dB SL for the NH groups) are shown in Fig. 3. The upper row shows results without vocoding; the lower row shows results with vocoding. The results from the young NH, age-matched NH, and HI listeners are shown in the left, middle and right columns, respectively.

Consider first the results from the non-vocoded conditions (upper row of Fig. 3). Performance in the presence of Gaussian noise (GN; filled squares) is quite similar across the three groups but, as expected, performance is slightly poorer in the HI group. When the noise masker is replaced by 16 pure tones (PT; open circles), performance is markedly better—all three groups show a large improvement in performance, although there appears to be a systematic decrease in performance from young NH to age-matched NH to the HI group. Imposing noise envelope fluctuations to the pure-tone maskers (MT; filled circles) results in a small reduction in performance for all three groups. It appears that both the spectral sparsity of the tonal maskers (PT and MT), as well as the lack of fluctuations (PT), contributes to the release from masking relative to the GN condition.

Fig. 3
figure 3

Speech intelligibility with the speech at roughly equal sensation level (SL) across the three groups. Sentence recognition was tested in speech-shaped Gaussian noise (GN), pure tones (PT), and modulated tones (MT)

Full size image

Consider next the results from the vocoded conditions (lower row of Fig. 3). Here, the benefits of frequency selectivity have been reduced by limiting spectral resolution to the 16 vocoder channels. As expected, the MT masker produces very similar results to the GN masker, as both produce very similar outputs from the tone vocoder. The young NH and the age-matched NH groups seem able to take similar advantage of the lack of inherent masker fluctuations in the PT condition. In contrast, the differences between the PT and MT seem less pronounced in the HI listeners; although some differences remain (in contrast to CI users), they are smaller than in the NH listeners.

The results using 85-dB SPL speech for all listeners are shown in Fig. 4. Only three speech-to-masker ratios (‑5, 0, and 5 dB) were tested in this part of the experiment. The relevant data from the HI listeners are simply replotted from Fig. 3. Performance was generally worse overall for both the young and age-matched NH listeners, compared to results at the lower sound level. In addition, the difference between the PT and MT conditions was reduced. This reduction was observed even in the non-vocoded conditions, but was particularly apparent in the vocoded conditions. In fact, in the vocoded conditions, the differences between the NH groups and the HI group were relatively small.

Fig. 4
figure 4

Speech intelligibility with the speech at equal sound pressure level (85 dB SPL). Data from the HI listeners are replotted from Fig. 3

Full size image

4 Discussion

Overall, the HI group showed smaller-than-normal differences between maskers with and without inherent fluctuations. This loss of sensitivity to inherent fluctuations was particularly apparent in the vocoded conditions (Fig. 3; lower right panel). However, in contrast to earlier results from CI users (Oxenham and Kreft 2014), some differences remained between conditions with and without inherent fluctuations, suggesting that the effects of poorer frequency selectivity are not as profound as for CI users. This could be for two reasons: First, frequency selectivity in HI listeners with mild-to-moderate hearing loss is not as poor as that in CI users. Second, the difference between the interaction occurring before and after envelope extraction may affect outcomes to some extent, in that the reduction in masker modulation power may be greater for CI users than for HI listeners, even with a similar loss of frequency selectivity.

One interesting outcome was that the differences between the NH groups and the HI group were not very pronounced when the speech was presented to the groups at the same high SPL. It is well known that frequency selectivity becomes poorer in NH listeners at high levels (e.g., Nelson and Freyman 1984; Glasberg and Moore 2000). Apparently the filter broadening with level in NH listeners is sufficient to reduce the effects of inherent masker fluctuations.

Overall, the results show that the importance of inherent masker fluctuations in determining speech intelligibility in noise depends to some extent on the conditions and the listening population. It cannot be claimed that inherent masker fluctuations always limit speech perception, as the effect of the fluctuations is non-existent in CI users (Oxenham and Kreft 2014), and is greatly reduced in HI listeners and in NH listeners at high sound levels. Finally, the reduced effect of inherent fluctuations may also provide a reason for why HI listeners (as well as CI users) exhibit less masking release in the presence of maskers with imposed additional fluctuations.