AJA
Research Article
High- and Low-Frequency Decrement Contribution to Spectral Enhancement Jeffrey J. DiGiovanni,a Travis L. Riffle,a and Naveen K. Nagarajb
Purpose: The relative benefit of a single, flanking, high- or low-frequency decrement was assessed to better understand properties of spectral enhancement that may aid in algorithm design. Method: Detection thresholds were measured for intensity increments applied to a narrow target band of frequencies embedded in a broadband signal while 400-Hz-wide, 9or 12-dB-deep intensity decrements were placed above, below, or on both sides of the target band. A mono condition with no decrements was used as a control. Eight participants with normal hearing and 8 participants with hearing impairment took part in this experiment.
Results: Performance improved in the presence of decrements for both groups and was equivalent for both high- or low-frequency decrements. Comparison with individually measured auditory filters revealed that participants with normal hearing made use of energy cues available within these filters, whereas some participants with hearing impairment, despite improved increment detection, underutilized this information. Conclusion: Inclusion of a single, adjacent, high- or lowfrequency decrement improved increment detection but not to the same extent as when decrements flanked both sides.
S
(NH) and listeners with hearing impairment (HI). It has been well established that, generally, the slope on the lowfrequency side of the auditory filter is steeper than the highfrequency side. This asymmetric shape of the auditory filter may be responsible for unequal contribution of the lowand high-frequency decrements. However, the extent to which individual high- and low-frequency decrements contributed to overall performance is unknown. Consequently, it can be deduced that energy on the high-frequency side would be attenuated less than that on the low-frequency side. For this reason, the asymmetry of auditory filters may reveal that one of the decrements may be dominating the observed improvements. If either high- or low-frequency decrements are found to be more beneficial than the other, it would be of interest to implement a one-sided decrement in cases in which flanking decrements would not be a feasible option (e.g., formant spacing is too close), especially in the development of an operational, real-time algorithm. To better understand the conditions that contribute to the benefits that spectral enhancement may provide and therefore ones that developers might use in an algorithm, 16 participants listened in an increment detection experiment. Conditions were created by adding a single decrement to only one side of the increment, adding decrements to both sides of the increment, or by completely removing the decrements from the spectrum. These conditions were designed to provide separate measures of increment detection performance for conditions with both high- and low-frequency
peech intelligibility has been identified as one of the most pervasive problems for hearing aid users with sensorineural hearing loss. Hearing aids provide amplification to compensate for shifted thresholds, but they fall short of enhancing clarity to improve speech intelligibility, especially in noise (Loizou & Kim, 2011). One method proposed to reduce distortion is spectral enhancement, whereby spectral peaks are amplified selectively within a signal while the valleys or troughs are attenuated or unaltered. Research evaluating efficacy of spectral enhancement strategies has resulted in mixed findings (Baer, Moore, & Gatehouse, 1993; DiGiovanni & Nair, 2006; DiGiovanni, Nelson, & Schlauch, 2005; Franck, van Kreveld-Bos, Dreschler, & Verschuure, 1999; Lyzenga, Festen, & Houtgast, 2002; Miller, Calhoun, & Young, 1999; Simpson, Moore, & Glasberg, 1990). DiGiovanni et al. (2005) and DiGiovanni and Nair (2006) used broadband stimuli with spectral decrements surrounding a spectral peak of narrowband noise to determine viability of spectral enhancement. When flanking spectral decrements were present, increment detection thresholds improved for both listeners with normal hearing
a
Ohio University, Athens UALR Speech and Hearing Clinic, Little Rock, AR
b
Correspondence to Jeffrey J. DiGiovanni: [email protected] Editor and Associate Editor: Larry Humes Received December 18, 2014 Revision received April 27, 2015 Accepted May 17, 2015 DOI: 10.1044/2015_AJA-14-0087
432
Disclosure: The authors have declared that no competing interests existed at the time of publication.
American Journal of Audiology • Vol. 24 • 432–439 • September 2015 • Copyright © 2015 American Speech-Language-Hearing Association
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
spectral decrements. Participants with HI or NH were required to detect a narrow-band peak (simulating a vowel formant) in a harmonic spectrum. Given the asymmetry of auditory filters, it was hypothesized that the decrement flanking the high-frequency side would contribute more to spectral increment detection performance than would the low-frequency decrement. These findings will aid in the development of a real-time, spectral-enhancement algorithm.
Method Participants Eight individuals (19–29 years of age; average = 25.1 years) with NH (thresholds of 15 dB HL at audiometric frequencies) and eight individuals (59–87 years of age; average = 73.9 years) with HI participated. The group with HI had mild-to-severe sensorineural hearing loss with thresholds ranging from 35 to 60 dB HL and from 1.5 to 3.0 kHz (see Table 1). Using Békésy discrete-frequency audiometry, all participants’ thresholds were verified to vary less than 10 dB between 1.5 and 2.5 kHz in 25-Hz steps. Auditory-filter bandwidths were measured on all participants using a modified method proposed by Glasberg and Moore (1990). Auditory bandwidths were measured at 2.0 kHz for all participants. Bandwidths were calculated as follows: W ðgÞ: ¼ ð1 rÞ ð1 þ pgÞepg þ r; where p determines the passband width and the filter skirt slope, and r is the dynamic range limiter. Auditory-filter bandwidths were derived by fitting the five data points into the two-parameter Roex ( p, r) model (Patterson, NimmoSmith, Weber, & Milroy, 1982). The average equivalent rectangular bandwidths (ERBs) of the participants with NH and HI were 268 and 945 Hz, respectively (see Table 2). In general, the average ERBs for both the groups with NH and HI were within the range found in the literature (Dubno & Dirks, 1989; Glasberg & Moore, 1986). The difference in performances of the groups with NH and HI can be attributed to the loss in frequency resolution. Table 1. Audiometric thresholds (dB HL) and ages (years) for all participants with hearing impairment (HI) are shown by frequency.
Age Participant (years) 0.25 HI 1 HI 2 HI 3 HI 4 HI 5 HI 6 HI 7 HI 8
69 82 87 61 70 78 85 59
25 60 35 35 25 20 45 15
Audiometric thresholds (kHz) 0.5
1.0
30 55 20 40 15 25 50 15
35 45 35 40 25 30 60 50
1.5
45 35 45
2.0
3.0
4.0
8.0
45 50 50 50 40 45 55 50
35
60 60 55 60 45 60 60 60
65 NR 85 55 45 NR 65 70
50
Note. NR = not reported.
Table 2. Ages and equivalent rectangular bandwidths (ERBs) for participants with hearing impairment (HI) and normal hearing (NH) along with group averages. Participant HI HI HI HI HI HI HI HI
Age (years)
ERB (Hz)
1 2 3 4 5 6 7 8
69 82 87 61 70 78 85 59
HI average
73.9
944.6
33 23 19 35 19 29 21
283.4 215.7 243.1 266.3 332.1 279.4 278.4 242.7
25.6
267.6
NH NH NH NH NH NH NH NH
1 2 3 4 5 6 7 8
NH average
416.5 1078.4 613.5 564.0 1980.0 1171.1 438.6 1294.9
Broadened auditory filters are a common finding in participants with HI (DiGiovanni & Nair, 2006; Dubno & Dirks, 1989; Glasberg & Moore, 1986). Overall, our participants with HI had an average ERB that was 3.52 times that of the average ERB of the participants with NH. These data were used to predict performance in increment detection thresholds.
Stimuli Stimuli consisted of a flat, harmonic spectrum (F0 = 50 Hz, band limited at 5000 Hz) 500 ms in duration and with a 500-ms interstimulus interval. The target stimulus for all conditions was a three harmonic (i.e., 100-Hz-wide) narrow-band increment centered at 2000 Hz—specifically, 1950, 2000, and 2050 Hz. To reduce the energy within a spectral region, 400-Hz-wide, 9-dB-deep spectral decrements were inserted adjacent to the increment in certain conditions. The high and low conditions included a decrement flanking only on either the high-frequency side (Hi) or the lowfrequency side (Lo) of the increment, respectively, and the bidecrement (Bi) condition included decrements placed on both sides of the increment. An additional mono condition containing no decrements served as the control. The convention used to refer to these conditions was the depth (dB)/width (Hz) for spectral location (mono, Lo, Hi, Bi). For instance, a 9-dB-depth, 400-Hz-wide decrement on both sides of the increment was referred to as 9/400Bi. Schematic spectra for the stimuli are shown in Figure 1. Stimuli were presented monaurally via earphones at 85 dB SPL.
Procedure A three-alternative, forced-choice task targeting 70.7% (two-down, one-up adaptive tracking) on the psychometric function (Levitt, 1971) was used to measure each of the
DiGiovanni et al.: Decrement Contribution
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
433
Figure 1. Spectra for the stimuli for the mono, bidecrement (Bi), low-frequency side of the increment (Lo), and high-frequency side of the increment (Hi) conditions are shown. Each line represents a harmonic. The dashed lines are the three harmonics that may be incremented (950, 1000, and 1050 Hz). The dotted lines represent an increment that was included in any target interval. The mono with the increment and the mono without the increment are shown in the top row, the 9/400Bi condition stimuli are shown in the middle row, and the 9/400Hi and 9/400Lo conditions are shown in the bottom row.
at threshold between the signal increment and the harmonic spectrum. The spectral enhancement conditions are thus expressed in terms of decibel increments above the spectrum level of the harmonic series. Benefit is the difference between the increment detection threshold in the mono condition and in a given decrement condition (e.g., mono and 9/400). Benefits for the groups with NH and HI are shown in Figure 2. Table 3 provides the mean detection threshold of the increment relative to that of the standard by group for each condition. Separate repeated-measures analyses of variance (ANOVAs) were completed for decrement depth and decrement location, as the mono condition served as a control for both of these variables. Decrement depth (mono, 9 dB, and 12 dB) or decrement location (mono, Lo, Hi, Bi) served as the within-factor variables for each main evaluation. Group (participants with NH or HI) served as a betweenfactor variable for both ANOVAs. The significance level for each ANOVA was halved (i.e., a = .025) to account for the multiple analyses. When depth served as the within factor, there was a significant main effect for depth, F(2, 28) = 24.42, p < .001, but not for group. There were no significant interactions. Post hoc pairwise comparisons (Fisher’s protected-t least significant difference multiple comparison test) were completed to explore the main effect of decrement depth. Results revealed that when compared with the mono condition, participants’ performances were significantly better when either a 9- or 12-dB decrement was present, but performance did not differ between decrement depths. When decrement location served as the within factor, significant effects were found for group, F(1, 14) = 7.09, p = .019, as well as decrement location, F(3, 42) = 14.55,
Figure 2. Group-mean improvements of the participants with normal hearing (NH) and hearing impairment (HI) are shown. Improvement was calculated by subtracting the group-mean average for a decrement condition from the mono condition. The upper panel shows these results for the conditions without the first formant (F1), and the lower panel shows these results for the conditions with the F1.
increment detection thresholds in the targeted spectral region within the harmonic series. Step size was 3 dB for the first two reversals, and it was 1 dB for the remaining reversals. A block of trials terminated once 10 reversals were obtained with the mean of the final eight reversals designated as threshold. The average threshold across three blocks was set as the participant’s threshold for each condition. Several conditions were used: mono, 9/400Bi, 12/400Bi, 9/400Lo, 12/400Lo, 9/400Hi, and 12/400Hi. The order of conditions was randomized among participants.
Results Increment thresholds were converted from 10 log10 (DI/I) to 10 log10 [DI/(I + 1)] to demonstrate the relative difference
434
American Journal of Audiology • Vol. 24 • 432–439 • September 2015
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
Table 3. Group-mean thresholds (dB) are shown by group and by condition, including the mono condition (i.e., 0 dB) as well as the lowdecrement, high-decrement, and bidecrement conditions. Group Depth of decrement 0 dB or mono 9 dB 9 dB 9 dB 12 dB 12 dB 12 dB
Decrement configuration
Participants with NH
Participants with HI
No decrement Low side High side Both sides Low side High side Both sides
5.0 (0.9) 3.5 (1.0) 3.8 (1.4) 3.3 (1.0) 3.8 (1.5) 3.6 (1.1) 3.0 (0.9)
8.1 (2.6) 6.6 (4.1) 7.9 (4.9) 6.9 (3.8) 7.4 (4.6) 7.3 (4.2) 6.8 (4.5)
Note. Standard deviations are shown in parentheses after each threshold. NH = normal hearing; HI = hearing impairment.
p < .001. There were no significant interactions. Post hoc pairwise comparisons (Fisher’s protected-t least significant difference multiple comparison test) for decrement type (mono, Lo, Hi, Bi) revealed that participants’ performance was significantly better for all decrement conditions compared with the mono condition. In addition, the performance in the Bi condition was significantly better than for the Hi and Lo conditions. There was not a significant difference in performance between Hi and Lo conditions. Therefore, participants performed better when at least one decrement was present regardless of its placement, and they benefited further with two present (i.e., decrement flanking).
Discussion Our outcomes suggest that this strategy of placing only a single decrement on either side of the increment provides benefit, albeit to a lesser degree than implementing both a high-frequency decrement and a low-frequency decrement. In the development of signal-processing algorithms, there may be situations in which a decrement may only be placed on one side of the increment because of the available spectra space—for example, high front vowels in which the first formant and the second formant are close in frequency and do not allow sufficient space for a decrement to be placed on the high side of the first formant. Therefore, it is a useful extension of previous findings that placing a decrement on either the low- or high-frequency side has demonstrated significant improvement. Although this study does not assess speech intelligibility measures directly, it has implications of furthering the development of hearing aid signal-processing algorithms that are designed to improve speech intelligibility. The findings of this study provide insight into potential signal-processing techniques that could be incorporated into novel algorithm designs. The contribution of single-sided decrements for increment detection may now be considered when discussing the benefits of spectral enhancement. On the basis of the asymmetric shape of the auditory filter, we hypothesized that the high-frequency decrements would be more beneficial than their low-frequency counterparts. This hypothesis was not verified, as either single-sided
decrement revealed similar benefit. It is possible that the difference in attenuation on the high- and low-filter skirts is too small to reveal a difference or that methods used were not sensitive enough to capture the effect. Finally, our results confirmed that inclusion of a decrement on both sides of this region provides even greater enhancement. Power within an auditory filter may be the determinant (Bilger, 1978; DiGiovanni & Nair, 2006; DiGiovanni et al., 2005), as our participants with HI had broader auditory filters and performed worse than the participants with NH. If the overall power within the auditory filter is the determining factor for performance, then the relative signal-to-noise ratio—or rather the increment-to-spectrum level ratio— would decrease with increasing bandwidth. Furthermore, this would explain the improvement measured for both groups when spectral decrements are inserted. Removing energy from the background spectrum results in an increase in the relative power of the increment to the spectrum. However, as the filter widens beyond the bandwidth of the decrement, the effect of removing energy from part of the spectrum becomes less significant and will ultimately result in reduced benefit. The overall-energy model was designed on the premise that the main cue for detection was the comparison of the overall energy in a given bandwidth (i.e., the auditoryfilter bandwidth) between the nondecrement condition and the decrement condition (DiGiovanni & Nair, 2006; DiGiovanni et al., 2005). Using this model, the predicted amount of improvement related to a given participant’s auditory-filter bandwidth can be calculated. The overall energy levels for the mono condition and each enhancement condition were calculated within each participant’s auditoryfilter bandwidth. To calculate the overall energy, the energy of each harmonic that fell within the auditory filter was summed. The predicted benefit for a particular condition was the difference in energy between the mono condition and the given enhancement condition. Tables 4 and 5 show the benefits predicted from the overall-energy model compared against actual improvements for the participants with NH and HI, respectively. To quantify the difference between the model output and participant performance, the root-mean-square (RMS)
DiGiovanni et al.: Decrement Contribution
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
435
436
American Journal of Audiology • Vol. 24 • 432–439 • September 2015
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx NH 2
NH 3
NH 4
NH 5
NH 6
NH 7
NH 8
1.2 1.4 1.2 1.4 2.6 3.1
1.2 0.8 1.6 1.7 2.2 2.5
1.2 1.4 1.2 1.4 2.6 3.1
2.3 1.9 0.7 0.8 2.3 2.9
1.2 1.4 1.2 1.4 2.6 3.1
1.6 1.6 1.2 1.6 2.5 3.1
1.2 1.4 1.2 1.4 2.6 3.1
1.1 0.6 0.2 −0.5 0.9 1.5
Note. Lo = low-frequency side of the increment; Hi = high-frequency side of the increment; Bi = bidecrement.
9/400Lo 12/400Lo 9/400Hi 12/400Hi 9/400Bi 12/400Bi
1.8 2.1 1.8 2.1 4.0 4.9
2.2 1.2 2.2 2.5 3.4 4.4
1.2 1.4 1.2 1.4 2.6 3.1
0.1 0.4 2.0 2.7 1.0 1.2
1.2 1.4 1.2 1.4 2.6 3.1
2.9 3.2 2.1 1.9 2.9 2.2
1.2 1.4 1.2 1.4 2.6 3.1
1.8 0.4 −0.4 0.9 0.7 1.1
Condition Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured
NH 1
Table 4. Predicted benefits using overall-energy model and measured benefit are shown by condition for each participant with normal hearing (NH) to allow direct comparisons between the model outcomes and measured thresholds.
DiGiovanni et al.: Decrement Contribution
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
437
HI 2
HI 3
HI 4
HI 5
HI 6
HI 7
HI 8
2.1 2.5 2.1 2.5 4.9 6.0
2.8 2.0 1.5 1.1 1.0 1.0
2.5 2.9 2.5 2.9 5.9 7.3
0.1 −2.9 −6.3 −2.7 −1.0 −0.9
2.5 3.0 2.5 3.0 6.0 7.5
3.8 2.2 0.2 −0.9 2.2 4.6
2.3 2.8 2.3 2.8 5.5 6.8
2.9 3.8 1.5 2.5 2.5 3.6
Note. Lo = low-frequency side of the increment; Hi = high-frequency side of the increment; Bi = bidecrement.
9/400Lo 12/400Lo 9/400Hi 12/400Hi 9/400Bi 12/400Bi
1.2 1.5 1.2 1.5 2.7 3.2
−0.7 −0.4 2.3 3.2 0.5 0.5
2.2 2.6 2.2 2.6 5.2 6.4
3.1 2.4 3.0 3.1 3.1 3.7
2.1 2.5 2.1 2.5 4.9 6.0
1.8 −1.7 −1.2 −1.3 0.9 −0.6
2.0 2.4 2.0 2.4 4.6 5.7
−0.2 0.5 0.3 1.4 0.6 −0.3
Condition Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured
HI 1
Table 5. Predicted benefits using overall-energy model and measured benefit are shown by condition for each participant with hearing impairment (HI) to allow direct comparisons between the model outcomes and measured thresholds.
was calculated, in dB, between the measured and predicted benefit. Table 6 displays the results of this analysis, both by group (separately and combined) and by condition. The mean RMS errors in the Bi conditions were 1.2 dB for the group with NH and 4.5 dB for the group with HI. When taking into account the average size of the auditory-filter bandwidth for the group with NH (267.6 Hz) and the group with HI (944.6 Hz), it is tempting to attribute the increased predictive error in the group with HI to the larger auditoryfilter bandwidth. However, it should be noted that the overall-energy model accounts for the auditory-filter bandwidth for each participant to calculate the predicted benefit. A possible explanation to partially account for the larger error in predictive accuracy of the group with HI is the decreasing contribution in the tails of the auditory filter. To calculate the predicted benefit, the ERB was used, for which each harmonic contributed equally regardless of its location relative to the center frequency. In cases of broader than normal auditory filters, the decremented harmonics fall within the tails of the filter and therefore would contribute less if an equivalent rectangular filter was used. If the frequencies closer to the center of the auditory filter contributed more to the measured benefit, then the model would overpredict as the auditory-filter bandwidths increases. In the Bi condition, the model overpredicted the benefit for every participant, resulting in an RMS error of 4.5 dB. However, the RMS error in the single-decrement conditions was only 2.5 dB. It is clear that the model overpredicts systematically for the Bi condition, but there is greater overlap in predicted and measured benefit in the single-sided decrement conditions, as is shown in Figure 3. Because ERBs were used to make predictions with the overall-energy model, it is possible that the consistent overprediction in the Bi conditions is a result of the energy in the tails being overrepresented because of the rectangular filter shape. To test this, data from participants with HI were used to generate individual auditory filters (Patterson et al., 1982) to replace the ERBs in the overall-energy model. When compared with the original predictions, the RMS error was 0.8 dB lower when using the more natural auditory-filter shape. Although this may address the Table 6. Root-mean-square (RMS) errors are shown by condition and group as well as for all participants combined. RMS error (dB) Condition
NH
HI
NH and HI
9/400Lo 12/400Lo 9/400Hi 12/400Hi 9/400Bi 12/400Bi
0.87 0.96 0.84 0.90 1.10 1.20
1.47 2.73 3.49 2.89 3.97 5.06
1.21 2.05 2.53 2.14 2.91 3.68
0.99
3.45
2.54
Combined
Note. NH = normal hearing; HI = hearing impairment; Lo = lowfrequency side of the increment; Hi = high-frequency side of the increment; Bi = bidecrement.
438
Figure 3. Listener benefits for individuals with HI (hearing impairment) are shown. The bidecrement (Bi) and one-sided-decrement conditions are plotted in the upper and lower panels, respectively. The overallenergy model’s predicted benefit is plotted for these conditions. The open symbols can be compared with the dotted line, and the closed symbols can be compared with the solid line. Lo = lowfrequency side of the increment; Hi = high-frequency side of the increment; ERB = equivalent rectangular bandwidth.
overprediction in part, there still remains >3.6 dB unaccounted for with this hypothesis. Despite numerous efforts to develop spectral enhancement algorithms, success in terms of improvement in speech intelligibility has been limited. However, data presented in this study support its viability. Every enhanced condition revealed improvements over the unenhanced condition. Moreover, the overall-energy model predicts with varying accuracy the benefits of the enhancements. It is important to note that we used simulated stimuli rather than an algorithm when collecting data in this study. Although this cannot be implemented into a device directly, a benefit of this approach is that any improvements would not be negated by improper choice of parameters required in an algorithm design. Earlier efforts focused on testing complete algorithms, rendering it difficult to ascertain whether spectral enhancement was a nonviable approach or whether the implementation was somehow flawed. The data presented here suggest that with careful implementation and further
American Journal of Audiology • Vol. 24 • 432–439 • September 2015
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
testing (e.g., speech testing in noise), spectral enhancement holds promise. However, other aspects of hearing aid processing may affect the success of spectral enhancement.
Conclusions Perceptual threshold improvements can be made by either selectively enhancing the peaks or attenuating the energy adjacent to the peak in the spectrum. Across conditions, participants with NH performed better in these tasks than participants with HI, even though benefit was measured for both groups. On the basis of the data presented, it can be concluded that removing energy surrounding an increment improves performance more than removing energy from only one side. Further, equal benefit can be expected regardless of the side from which the energy is removed. Although removing energy from one side results in less improvement than removing it from both sides, it is still worthwhile to remove energy from one side. This may be more practical given signal-processing limitations and because, in certain conditions, formants may be too closely spaced to remove sufficient energy between them. These findings may be useful in defining the parameters for a spectral-enhancement algorithm.
Acknowledgment This study was funded by an American Speech-LanguageHearing Foundation Research Grant awarded to Jeffrey J. DiGiovanni.
References Baer, T., Moore, B. C., & Gatehouse, S. (1993). Spectral contrast enhancement of speech in noise for listeners with sensorineural hearing impairment: Effects of intelligibility, quality, and response times. Journal of Rehabilitation Research and Development, 30, 49–72. Bilger, R. C. (1978). A revised critical band hypothesis. In S. K. Hirsh, D. H. Eldridge, I. J. Hirsh, & S. R. Silverman (Eds.),
Hearing and Davis: Essays honoring Hallowell Davis (pp. 191–198). St. Louis, MO: Washington University Press. DiGiovanni, J. J., & Nair, P. (2006). Auditory filters and the benefit measured from spectral enhancement. The Journal of the Acoustical Society of America, 120, 1529–1538. DiGiovanni, J. J., Nelson, P. B., & Schlauch, R. S. (2005). A psychoacoustic evaluation of spectral enhancement. Journal of Speech, Language, and Hearing Research, 48, 1–15. Dubno, J. R., & Dirks, D. D. (1989). Auditory filter characteristics and consonant recognition for hearing-impaired listeners. The Journal of the Acoustical Society of America, 85, 1666–1675. Franck, B. A., van Kreveld-Bos, C. S., Dreschler, W. A., & Verschuure, H. (1999). Evaluation of spectral enhancement in hearing aids, combined with phonemic compression. The Journal of the Acoustical Society of America, 106, 1452–1464. Glasberg, B. R., & Moore, B. C. (1986). Auditory filter shapes in subjects with unilateral and bilateral cochlear impairments. The Journal of the Acoustical Society of America, 79, 1020–1033. Glasberg, B. R., & Moore, B. C. (1990). Deviation of auditory filter shapes from notch-noise data. Hearing Research, 47, 103–138. Levitt, H. (1971). Transformed up-down methods in psychoacoustics. The Journal of the Acoustical Society of America, 49, 467–477. Loizou, P. C., & Kim, G. (2011). Gain-induced speech distortions and the absence of intelligibility benefit with existing noisereduction algorithms. The Journal of the Acoustical Society of America, 130, 1581–1596. Lyzenga, J., Festen, J. M., & Houtgast, T. (2002). A speech enhancement scheme incorporating spectral expansion evaluated with simulated loss of frequency selectivity. The Journal of the Acoustical Society of America, 112, 1145–1157. Miller, R. L., Calhoun, B. M., & Young, E. D. (1999). Contrast enhancement improves the representation of /ɛ/-like vowels in the hearing-impaired auditory nerve. The Journal of the Acoustical Society of America, 106, 2693–2708. Patterson, R. D., Nimmo-Smith, I., Weber, D. L., & Milroy, R. (1982). The deterioration of hearing with age: Frequency selectivity, the critical ratio, the audiogram, and speech threshold. The Journal of the Acoustical Society of America, 72, 1788–1803. Simpson, A. M., Moore, B. C., & Glasberg, B. R. (1990). Spectral enhancement to improve the intelligibility of speech in noise for hearing-impaired listeners. Acta Otolaryngologica, 469, 101–107.
DiGiovanni et al.: Decrement Contribution
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
439
Research Article
High- and Low-Frequency Decrement Contribution to Spectral Enhancement Jeffrey J. DiGiovanni,a Travis L. Riffle,a and Naveen K. Nagarajb
Purpose: The relative benefit of a single, flanking, high- or low-frequency decrement was assessed to better understand properties of spectral enhancement that may aid in algorithm design. Method: Detection thresholds were measured for intensity increments applied to a narrow target band of frequencies embedded in a broadband signal while 400-Hz-wide, 9or 12-dB-deep intensity decrements were placed above, below, or on both sides of the target band. A mono condition with no decrements was used as a control. Eight participants with normal hearing and 8 participants with hearing impairment took part in this experiment.
Results: Performance improved in the presence of decrements for both groups and was equivalent for both high- or low-frequency decrements. Comparison with individually measured auditory filters revealed that participants with normal hearing made use of energy cues available within these filters, whereas some participants with hearing impairment, despite improved increment detection, underutilized this information. Conclusion: Inclusion of a single, adjacent, high- or lowfrequency decrement improved increment detection but not to the same extent as when decrements flanked both sides.
S
(NH) and listeners with hearing impairment (HI). It has been well established that, generally, the slope on the lowfrequency side of the auditory filter is steeper than the highfrequency side. This asymmetric shape of the auditory filter may be responsible for unequal contribution of the lowand high-frequency decrements. However, the extent to which individual high- and low-frequency decrements contributed to overall performance is unknown. Consequently, it can be deduced that energy on the high-frequency side would be attenuated less than that on the low-frequency side. For this reason, the asymmetry of auditory filters may reveal that one of the decrements may be dominating the observed improvements. If either high- or low-frequency decrements are found to be more beneficial than the other, it would be of interest to implement a one-sided decrement in cases in which flanking decrements would not be a feasible option (e.g., formant spacing is too close), especially in the development of an operational, real-time algorithm. To better understand the conditions that contribute to the benefits that spectral enhancement may provide and therefore ones that developers might use in an algorithm, 16 participants listened in an increment detection experiment. Conditions were created by adding a single decrement to only one side of the increment, adding decrements to both sides of the increment, or by completely removing the decrements from the spectrum. These conditions were designed to provide separate measures of increment detection performance for conditions with both high- and low-frequency
peech intelligibility has been identified as one of the most pervasive problems for hearing aid users with sensorineural hearing loss. Hearing aids provide amplification to compensate for shifted thresholds, but they fall short of enhancing clarity to improve speech intelligibility, especially in noise (Loizou & Kim, 2011). One method proposed to reduce distortion is spectral enhancement, whereby spectral peaks are amplified selectively within a signal while the valleys or troughs are attenuated or unaltered. Research evaluating efficacy of spectral enhancement strategies has resulted in mixed findings (Baer, Moore, & Gatehouse, 1993; DiGiovanni & Nair, 2006; DiGiovanni, Nelson, & Schlauch, 2005; Franck, van Kreveld-Bos, Dreschler, & Verschuure, 1999; Lyzenga, Festen, & Houtgast, 2002; Miller, Calhoun, & Young, 1999; Simpson, Moore, & Glasberg, 1990). DiGiovanni et al. (2005) and DiGiovanni and Nair (2006) used broadband stimuli with spectral decrements surrounding a spectral peak of narrowband noise to determine viability of spectral enhancement. When flanking spectral decrements were present, increment detection thresholds improved for both listeners with normal hearing
a
Ohio University, Athens UALR Speech and Hearing Clinic, Little Rock, AR
b
Correspondence to Jeffrey J. DiGiovanni: [email protected] Editor and Associate Editor: Larry Humes Received December 18, 2014 Revision received April 27, 2015 Accepted May 17, 2015 DOI: 10.1044/2015_AJA-14-0087
432
Disclosure: The authors have declared that no competing interests existed at the time of publication.
American Journal of Audiology • Vol. 24 • 432–439 • September 2015 • Copyright © 2015 American Speech-Language-Hearing Association
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
spectral decrements. Participants with HI or NH were required to detect a narrow-band peak (simulating a vowel formant) in a harmonic spectrum. Given the asymmetry of auditory filters, it was hypothesized that the decrement flanking the high-frequency side would contribute more to spectral increment detection performance than would the low-frequency decrement. These findings will aid in the development of a real-time, spectral-enhancement algorithm.
Method Participants Eight individuals (19–29 years of age; average = 25.1 years) with NH (thresholds of 15 dB HL at audiometric frequencies) and eight individuals (59–87 years of age; average = 73.9 years) with HI participated. The group with HI had mild-to-severe sensorineural hearing loss with thresholds ranging from 35 to 60 dB HL and from 1.5 to 3.0 kHz (see Table 1). Using Békésy discrete-frequency audiometry, all participants’ thresholds were verified to vary less than 10 dB between 1.5 and 2.5 kHz in 25-Hz steps. Auditory-filter bandwidths were measured on all participants using a modified method proposed by Glasberg and Moore (1990). Auditory bandwidths were measured at 2.0 kHz for all participants. Bandwidths were calculated as follows: W ðgÞ: ¼ ð1 rÞ ð1 þ pgÞepg þ r; where p determines the passband width and the filter skirt slope, and r is the dynamic range limiter. Auditory-filter bandwidths were derived by fitting the five data points into the two-parameter Roex ( p, r) model (Patterson, NimmoSmith, Weber, & Milroy, 1982). The average equivalent rectangular bandwidths (ERBs) of the participants with NH and HI were 268 and 945 Hz, respectively (see Table 2). In general, the average ERBs for both the groups with NH and HI were within the range found in the literature (Dubno & Dirks, 1989; Glasberg & Moore, 1986). The difference in performances of the groups with NH and HI can be attributed to the loss in frequency resolution. Table 1. Audiometric thresholds (dB HL) and ages (years) for all participants with hearing impairment (HI) are shown by frequency.
Age Participant (years) 0.25 HI 1 HI 2 HI 3 HI 4 HI 5 HI 6 HI 7 HI 8
69 82 87 61 70 78 85 59
25 60 35 35 25 20 45 15
Audiometric thresholds (kHz) 0.5
1.0
30 55 20 40 15 25 50 15
35 45 35 40 25 30 60 50
1.5
45 35 45
2.0
3.0
4.0
8.0
45 50 50 50 40 45 55 50
35
60 60 55 60 45 60 60 60
65 NR 85 55 45 NR 65 70
50
Note. NR = not reported.
Table 2. Ages and equivalent rectangular bandwidths (ERBs) for participants with hearing impairment (HI) and normal hearing (NH) along with group averages. Participant HI HI HI HI HI HI HI HI
Age (years)
ERB (Hz)
1 2 3 4 5 6 7 8
69 82 87 61 70 78 85 59
HI average
73.9
944.6
33 23 19 35 19 29 21
283.4 215.7 243.1 266.3 332.1 279.4 278.4 242.7
25.6
267.6
NH NH NH NH NH NH NH NH
1 2 3 4 5 6 7 8
NH average
416.5 1078.4 613.5 564.0 1980.0 1171.1 438.6 1294.9
Broadened auditory filters are a common finding in participants with HI (DiGiovanni & Nair, 2006; Dubno & Dirks, 1989; Glasberg & Moore, 1986). Overall, our participants with HI had an average ERB that was 3.52 times that of the average ERB of the participants with NH. These data were used to predict performance in increment detection thresholds.
Stimuli Stimuli consisted of a flat, harmonic spectrum (F0 = 50 Hz, band limited at 5000 Hz) 500 ms in duration and with a 500-ms interstimulus interval. The target stimulus for all conditions was a three harmonic (i.e., 100-Hz-wide) narrow-band increment centered at 2000 Hz—specifically, 1950, 2000, and 2050 Hz. To reduce the energy within a spectral region, 400-Hz-wide, 9-dB-deep spectral decrements were inserted adjacent to the increment in certain conditions. The high and low conditions included a decrement flanking only on either the high-frequency side (Hi) or the lowfrequency side (Lo) of the increment, respectively, and the bidecrement (Bi) condition included decrements placed on both sides of the increment. An additional mono condition containing no decrements served as the control. The convention used to refer to these conditions was the depth (dB)/width (Hz) for spectral location (mono, Lo, Hi, Bi). For instance, a 9-dB-depth, 400-Hz-wide decrement on both sides of the increment was referred to as 9/400Bi. Schematic spectra for the stimuli are shown in Figure 1. Stimuli were presented monaurally via earphones at 85 dB SPL.
Procedure A three-alternative, forced-choice task targeting 70.7% (two-down, one-up adaptive tracking) on the psychometric function (Levitt, 1971) was used to measure each of the
DiGiovanni et al.: Decrement Contribution
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
433
Figure 1. Spectra for the stimuli for the mono, bidecrement (Bi), low-frequency side of the increment (Lo), and high-frequency side of the increment (Hi) conditions are shown. Each line represents a harmonic. The dashed lines are the three harmonics that may be incremented (950, 1000, and 1050 Hz). The dotted lines represent an increment that was included in any target interval. The mono with the increment and the mono without the increment are shown in the top row, the 9/400Bi condition stimuli are shown in the middle row, and the 9/400Hi and 9/400Lo conditions are shown in the bottom row.
at threshold between the signal increment and the harmonic spectrum. The spectral enhancement conditions are thus expressed in terms of decibel increments above the spectrum level of the harmonic series. Benefit is the difference between the increment detection threshold in the mono condition and in a given decrement condition (e.g., mono and 9/400). Benefits for the groups with NH and HI are shown in Figure 2. Table 3 provides the mean detection threshold of the increment relative to that of the standard by group for each condition. Separate repeated-measures analyses of variance (ANOVAs) were completed for decrement depth and decrement location, as the mono condition served as a control for both of these variables. Decrement depth (mono, 9 dB, and 12 dB) or decrement location (mono, Lo, Hi, Bi) served as the within-factor variables for each main evaluation. Group (participants with NH or HI) served as a betweenfactor variable for both ANOVAs. The significance level for each ANOVA was halved (i.e., a = .025) to account for the multiple analyses. When depth served as the within factor, there was a significant main effect for depth, F(2, 28) = 24.42, p < .001, but not for group. There were no significant interactions. Post hoc pairwise comparisons (Fisher’s protected-t least significant difference multiple comparison test) were completed to explore the main effect of decrement depth. Results revealed that when compared with the mono condition, participants’ performances were significantly better when either a 9- or 12-dB decrement was present, but performance did not differ between decrement depths. When decrement location served as the within factor, significant effects were found for group, F(1, 14) = 7.09, p = .019, as well as decrement location, F(3, 42) = 14.55,
Figure 2. Group-mean improvements of the participants with normal hearing (NH) and hearing impairment (HI) are shown. Improvement was calculated by subtracting the group-mean average for a decrement condition from the mono condition. The upper panel shows these results for the conditions without the first formant (F1), and the lower panel shows these results for the conditions with the F1.
increment detection thresholds in the targeted spectral region within the harmonic series. Step size was 3 dB for the first two reversals, and it was 1 dB for the remaining reversals. A block of trials terminated once 10 reversals were obtained with the mean of the final eight reversals designated as threshold. The average threshold across three blocks was set as the participant’s threshold for each condition. Several conditions were used: mono, 9/400Bi, 12/400Bi, 9/400Lo, 12/400Lo, 9/400Hi, and 12/400Hi. The order of conditions was randomized among participants.
Results Increment thresholds were converted from 10 log10 (DI/I) to 10 log10 [DI/(I + 1)] to demonstrate the relative difference
434
American Journal of Audiology • Vol. 24 • 432–439 • September 2015
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
Table 3. Group-mean thresholds (dB) are shown by group and by condition, including the mono condition (i.e., 0 dB) as well as the lowdecrement, high-decrement, and bidecrement conditions. Group Depth of decrement 0 dB or mono 9 dB 9 dB 9 dB 12 dB 12 dB 12 dB
Decrement configuration
Participants with NH
Participants with HI
No decrement Low side High side Both sides Low side High side Both sides
5.0 (0.9) 3.5 (1.0) 3.8 (1.4) 3.3 (1.0) 3.8 (1.5) 3.6 (1.1) 3.0 (0.9)
8.1 (2.6) 6.6 (4.1) 7.9 (4.9) 6.9 (3.8) 7.4 (4.6) 7.3 (4.2) 6.8 (4.5)
Note. Standard deviations are shown in parentheses after each threshold. NH = normal hearing; HI = hearing impairment.
p < .001. There were no significant interactions. Post hoc pairwise comparisons (Fisher’s protected-t least significant difference multiple comparison test) for decrement type (mono, Lo, Hi, Bi) revealed that participants’ performance was significantly better for all decrement conditions compared with the mono condition. In addition, the performance in the Bi condition was significantly better than for the Hi and Lo conditions. There was not a significant difference in performance between Hi and Lo conditions. Therefore, participants performed better when at least one decrement was present regardless of its placement, and they benefited further with two present (i.e., decrement flanking).
Discussion Our outcomes suggest that this strategy of placing only a single decrement on either side of the increment provides benefit, albeit to a lesser degree than implementing both a high-frequency decrement and a low-frequency decrement. In the development of signal-processing algorithms, there may be situations in which a decrement may only be placed on one side of the increment because of the available spectra space—for example, high front vowels in which the first formant and the second formant are close in frequency and do not allow sufficient space for a decrement to be placed on the high side of the first formant. Therefore, it is a useful extension of previous findings that placing a decrement on either the low- or high-frequency side has demonstrated significant improvement. Although this study does not assess speech intelligibility measures directly, it has implications of furthering the development of hearing aid signal-processing algorithms that are designed to improve speech intelligibility. The findings of this study provide insight into potential signal-processing techniques that could be incorporated into novel algorithm designs. The contribution of single-sided decrements for increment detection may now be considered when discussing the benefits of spectral enhancement. On the basis of the asymmetric shape of the auditory filter, we hypothesized that the high-frequency decrements would be more beneficial than their low-frequency counterparts. This hypothesis was not verified, as either single-sided
decrement revealed similar benefit. It is possible that the difference in attenuation on the high- and low-filter skirts is too small to reveal a difference or that methods used were not sensitive enough to capture the effect. Finally, our results confirmed that inclusion of a decrement on both sides of this region provides even greater enhancement. Power within an auditory filter may be the determinant (Bilger, 1978; DiGiovanni & Nair, 2006; DiGiovanni et al., 2005), as our participants with HI had broader auditory filters and performed worse than the participants with NH. If the overall power within the auditory filter is the determining factor for performance, then the relative signal-to-noise ratio—or rather the increment-to-spectrum level ratio— would decrease with increasing bandwidth. Furthermore, this would explain the improvement measured for both groups when spectral decrements are inserted. Removing energy from the background spectrum results in an increase in the relative power of the increment to the spectrum. However, as the filter widens beyond the bandwidth of the decrement, the effect of removing energy from part of the spectrum becomes less significant and will ultimately result in reduced benefit. The overall-energy model was designed on the premise that the main cue for detection was the comparison of the overall energy in a given bandwidth (i.e., the auditoryfilter bandwidth) between the nondecrement condition and the decrement condition (DiGiovanni & Nair, 2006; DiGiovanni et al., 2005). Using this model, the predicted amount of improvement related to a given participant’s auditory-filter bandwidth can be calculated. The overall energy levels for the mono condition and each enhancement condition were calculated within each participant’s auditoryfilter bandwidth. To calculate the overall energy, the energy of each harmonic that fell within the auditory filter was summed. The predicted benefit for a particular condition was the difference in energy between the mono condition and the given enhancement condition. Tables 4 and 5 show the benefits predicted from the overall-energy model compared against actual improvements for the participants with NH and HI, respectively. To quantify the difference between the model output and participant performance, the root-mean-square (RMS)
DiGiovanni et al.: Decrement Contribution
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
435
436
American Journal of Audiology • Vol. 24 • 432–439 • September 2015
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx NH 2
NH 3
NH 4
NH 5
NH 6
NH 7
NH 8
1.2 1.4 1.2 1.4 2.6 3.1
1.2 0.8 1.6 1.7 2.2 2.5
1.2 1.4 1.2 1.4 2.6 3.1
2.3 1.9 0.7 0.8 2.3 2.9
1.2 1.4 1.2 1.4 2.6 3.1
1.6 1.6 1.2 1.6 2.5 3.1
1.2 1.4 1.2 1.4 2.6 3.1
1.1 0.6 0.2 −0.5 0.9 1.5
Note. Lo = low-frequency side of the increment; Hi = high-frequency side of the increment; Bi = bidecrement.
9/400Lo 12/400Lo 9/400Hi 12/400Hi 9/400Bi 12/400Bi
1.8 2.1 1.8 2.1 4.0 4.9
2.2 1.2 2.2 2.5 3.4 4.4
1.2 1.4 1.2 1.4 2.6 3.1
0.1 0.4 2.0 2.7 1.0 1.2
1.2 1.4 1.2 1.4 2.6 3.1
2.9 3.2 2.1 1.9 2.9 2.2
1.2 1.4 1.2 1.4 2.6 3.1
1.8 0.4 −0.4 0.9 0.7 1.1
Condition Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured
NH 1
Table 4. Predicted benefits using overall-energy model and measured benefit are shown by condition for each participant with normal hearing (NH) to allow direct comparisons between the model outcomes and measured thresholds.
DiGiovanni et al.: Decrement Contribution
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
437
HI 2
HI 3
HI 4
HI 5
HI 6
HI 7
HI 8
2.1 2.5 2.1 2.5 4.9 6.0
2.8 2.0 1.5 1.1 1.0 1.0
2.5 2.9 2.5 2.9 5.9 7.3
0.1 −2.9 −6.3 −2.7 −1.0 −0.9
2.5 3.0 2.5 3.0 6.0 7.5
3.8 2.2 0.2 −0.9 2.2 4.6
2.3 2.8 2.3 2.8 5.5 6.8
2.9 3.8 1.5 2.5 2.5 3.6
Note. Lo = low-frequency side of the increment; Hi = high-frequency side of the increment; Bi = bidecrement.
9/400Lo 12/400Lo 9/400Hi 12/400Hi 9/400Bi 12/400Bi
1.2 1.5 1.2 1.5 2.7 3.2
−0.7 −0.4 2.3 3.2 0.5 0.5
2.2 2.6 2.2 2.6 5.2 6.4
3.1 2.4 3.0 3.1 3.1 3.7
2.1 2.5 2.1 2.5 4.9 6.0
1.8 −1.7 −1.2 −1.3 0.9 −0.6
2.0 2.4 2.0 2.4 4.6 5.7
−0.2 0.5 0.3 1.4 0.6 −0.3
Condition Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured Predicted Measured
HI 1
Table 5. Predicted benefits using overall-energy model and measured benefit are shown by condition for each participant with hearing impairment (HI) to allow direct comparisons between the model outcomes and measured thresholds.
was calculated, in dB, between the measured and predicted benefit. Table 6 displays the results of this analysis, both by group (separately and combined) and by condition. The mean RMS errors in the Bi conditions were 1.2 dB for the group with NH and 4.5 dB for the group with HI. When taking into account the average size of the auditory-filter bandwidth for the group with NH (267.6 Hz) and the group with HI (944.6 Hz), it is tempting to attribute the increased predictive error in the group with HI to the larger auditoryfilter bandwidth. However, it should be noted that the overall-energy model accounts for the auditory-filter bandwidth for each participant to calculate the predicted benefit. A possible explanation to partially account for the larger error in predictive accuracy of the group with HI is the decreasing contribution in the tails of the auditory filter. To calculate the predicted benefit, the ERB was used, for which each harmonic contributed equally regardless of its location relative to the center frequency. In cases of broader than normal auditory filters, the decremented harmonics fall within the tails of the filter and therefore would contribute less if an equivalent rectangular filter was used. If the frequencies closer to the center of the auditory filter contributed more to the measured benefit, then the model would overpredict as the auditory-filter bandwidths increases. In the Bi condition, the model overpredicted the benefit for every participant, resulting in an RMS error of 4.5 dB. However, the RMS error in the single-decrement conditions was only 2.5 dB. It is clear that the model overpredicts systematically for the Bi condition, but there is greater overlap in predicted and measured benefit in the single-sided decrement conditions, as is shown in Figure 3. Because ERBs were used to make predictions with the overall-energy model, it is possible that the consistent overprediction in the Bi conditions is a result of the energy in the tails being overrepresented because of the rectangular filter shape. To test this, data from participants with HI were used to generate individual auditory filters (Patterson et al., 1982) to replace the ERBs in the overall-energy model. When compared with the original predictions, the RMS error was 0.8 dB lower when using the more natural auditory-filter shape. Although this may address the Table 6. Root-mean-square (RMS) errors are shown by condition and group as well as for all participants combined. RMS error (dB) Condition
NH
HI
NH and HI
9/400Lo 12/400Lo 9/400Hi 12/400Hi 9/400Bi 12/400Bi
0.87 0.96 0.84 0.90 1.10 1.20
1.47 2.73 3.49 2.89 3.97 5.06
1.21 2.05 2.53 2.14 2.91 3.68
0.99
3.45
2.54
Combined
Note. NH = normal hearing; HI = hearing impairment; Lo = lowfrequency side of the increment; Hi = high-frequency side of the increment; Bi = bidecrement.
438
Figure 3. Listener benefits for individuals with HI (hearing impairment) are shown. The bidecrement (Bi) and one-sided-decrement conditions are plotted in the upper and lower panels, respectively. The overallenergy model’s predicted benefit is plotted for these conditions. The open symbols can be compared with the dotted line, and the closed symbols can be compared with the solid line. Lo = lowfrequency side of the increment; Hi = high-frequency side of the increment; ERB = equivalent rectangular bandwidth.
overprediction in part, there still remains >3.6 dB unaccounted for with this hypothesis. Despite numerous efforts to develop spectral enhancement algorithms, success in terms of improvement in speech intelligibility has been limited. However, data presented in this study support its viability. Every enhanced condition revealed improvements over the unenhanced condition. Moreover, the overall-energy model predicts with varying accuracy the benefits of the enhancements. It is important to note that we used simulated stimuli rather than an algorithm when collecting data in this study. Although this cannot be implemented into a device directly, a benefit of this approach is that any improvements would not be negated by improper choice of parameters required in an algorithm design. Earlier efforts focused on testing complete algorithms, rendering it difficult to ascertain whether spectral enhancement was a nonviable approach or whether the implementation was somehow flawed. The data presented here suggest that with careful implementation and further
American Journal of Audiology • Vol. 24 • 432–439 • September 2015
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
testing (e.g., speech testing in noise), spectral enhancement holds promise. However, other aspects of hearing aid processing may affect the success of spectral enhancement.
Conclusions Perceptual threshold improvements can be made by either selectively enhancing the peaks or attenuating the energy adjacent to the peak in the spectrum. Across conditions, participants with NH performed better in these tasks than participants with HI, even though benefit was measured for both groups. On the basis of the data presented, it can be concluded that removing energy surrounding an increment improves performance more than removing energy from only one side. Further, equal benefit can be expected regardless of the side from which the energy is removed. Although removing energy from one side results in less improvement than removing it from both sides, it is still worthwhile to remove energy from one side. This may be more practical given signal-processing limitations and because, in certain conditions, formants may be too closely spaced to remove sufficient energy between them. These findings may be useful in defining the parameters for a spectral-enhancement algorithm.
Acknowledgment This study was funded by an American Speech-LanguageHearing Foundation Research Grant awarded to Jeffrey J. DiGiovanni.
References Baer, T., Moore, B. C., & Gatehouse, S. (1993). Spectral contrast enhancement of speech in noise for listeners with sensorineural hearing impairment: Effects of intelligibility, quality, and response times. Journal of Rehabilitation Research and Development, 30, 49–72. Bilger, R. C. (1978). A revised critical band hypothesis. In S. K. Hirsh, D. H. Eldridge, I. J. Hirsh, & S. R. Silverman (Eds.),
Hearing and Davis: Essays honoring Hallowell Davis (pp. 191–198). St. Louis, MO: Washington University Press. DiGiovanni, J. J., & Nair, P. (2006). Auditory filters and the benefit measured from spectral enhancement. The Journal of the Acoustical Society of America, 120, 1529–1538. DiGiovanni, J. J., Nelson, P. B., & Schlauch, R. S. (2005). A psychoacoustic evaluation of spectral enhancement. Journal of Speech, Language, and Hearing Research, 48, 1–15. Dubno, J. R., & Dirks, D. D. (1989). Auditory filter characteristics and consonant recognition for hearing-impaired listeners. The Journal of the Acoustical Society of America, 85, 1666–1675. Franck, B. A., van Kreveld-Bos, C. S., Dreschler, W. A., & Verschuure, H. (1999). Evaluation of spectral enhancement in hearing aids, combined with phonemic compression. The Journal of the Acoustical Society of America, 106, 1452–1464. Glasberg, B. R., & Moore, B. C. (1986). Auditory filter shapes in subjects with unilateral and bilateral cochlear impairments. The Journal of the Acoustical Society of America, 79, 1020–1033. Glasberg, B. R., & Moore, B. C. (1990). Deviation of auditory filter shapes from notch-noise data. Hearing Research, 47, 103–138. Levitt, H. (1971). Transformed up-down methods in psychoacoustics. The Journal of the Acoustical Society of America, 49, 467–477. Loizou, P. C., & Kim, G. (2011). Gain-induced speech distortions and the absence of intelligibility benefit with existing noisereduction algorithms. The Journal of the Acoustical Society of America, 130, 1581–1596. Lyzenga, J., Festen, J. M., & Houtgast, T. (2002). A speech enhancement scheme incorporating spectral expansion evaluated with simulated loss of frequency selectivity. The Journal of the Acoustical Society of America, 112, 1145–1157. Miller, R. L., Calhoun, B. M., & Young, E. D. (1999). Contrast enhancement improves the representation of /ɛ/-like vowels in the hearing-impaired auditory nerve. The Journal of the Acoustical Society of America, 106, 2693–2708. Patterson, R. D., Nimmo-Smith, I., Weber, D. L., & Milroy, R. (1982). The deterioration of hearing with age: Frequency selectivity, the critical ratio, the audiogram, and speech threshold. The Journal of the Acoustical Society of America, 72, 1788–1803. Simpson, A. M., Moore, B. C., & Glasberg, B. R. (1990). Spectral enhancement to improve the intelligibility of speech in noise for hearing-impaired listeners. Acta Otolaryngologica, 469, 101–107.
DiGiovanni et al.: Decrement Contribution
Downloaded From: http://aja.pubs.asha.org/ by a New York University User on 02/01/2016 Terms of Use: http://pubs.asha.org/ss/rights_and_permissions.aspx
439
