Preferred Compression Speed for Speech and Music and Its Relationship to Sensitivity to Temporal Fine Structure

Subjects

Twenty-two subjects (14 men) with moderate-to-severe sensorineural hearing loss were paid to participate. The number of subjects was chosen such that a correlation as small as .36 between a measure of TFS sensitivity and preference scores would be significant based on a one-tailed test. Their ages ranged from 56 to 87 years, with a mean of 73 years. Twenty were current users of multichannel compression hearing aids, and two did not use hearing aids. Audiometric thresholds were measured for all audiometric frequencies from 0.25 to 10 kHz, using a Grason-Stadler GSI61 audiometer (Eaden Prairie, MN), following the procedure recommended by the British Society of Audiology (2004). Telephonics TDH50 headphones (Farmington, NY) were used for audiometric frequencies up to 8 kHz, and Sennheiser HDA200 headphones (Wedemark, Germany) were used for the frequency of 10 kHz. Only the better ear of each subject was tested using the paired-comparison procedure. The mean and individual audiograms for the ears used for the preference judgments are shown in Figure 1. The hearing loss in the test ear ranged from 8 to 60 dB at 500 Hz, 6 to 64 dB at 1000 Hz, 26 to 70 dB at 2 kHz, and 48 to 74 dB at 4000 Hz. Eighteen subjects were native speakers of British English, and four were native speakers of Polish. OPEN IN VIEWER

Simulated Hearing Aid

The simulated hearing aid was the same as described by Moore, Füllgrabe, and Stone (2010) and Moore and Sek (2013). Briefly, the aid included a digital filter for overall shaping of the frequency response prior to splitting the signal into five channels, with independent compression in each channel. The lower and upper edge frequencies of channels 1 to 5, respectively, were 100 and 707 Hz, 707 and 1414 Hz, 1414 and 2828 Hz, 2828 and 5657 Hz, and 5657 and 10574 Hz. Five channels were used to achieve a compromise between having sufficient flexibility to implement the frequency-dependent insertion gains (IGs) and CRs required to match the prescription for each subject while avoiding the reduction of spectral and temporal contrast that can occur when many channels are used (Bor, Souza, & Wright, 2008; Plomp, 1988; Stone & Moore, 2008).

The IGs for a 65-dB speech-shaped noise and the CRs for the five channels were set according to the CAM2 prescription method (Moore, Glasberg, et al., 2010), modified slightly as described in Moore and Sek (2013, 2016). The version of the software used, called CAM2A, was 1.0.7.0. The compression thresholds were chosen to be representative of those used by default in the fitting software of some hearing-aid manufacturers. The compression thresholds (CTs) were set to 49, 41, 40, 34, and 28 dB SPL for channels 1 to 5, respectively.

To simulate fast compression, the attach and release times were set to 10 and 100 ms, respectively, for all channels. To simulate slow compression, the attack and release times were set to 50 and 3000 ms, respectively, for all channels. The CR was limited to 3 when fast compression was used because there is evidence that with fast compression high CRs can lead to reduced speech intelligibility (Verschuure et al., 1996). The CR was allowed to have any value up to 10 when slow compression was used. Table 1 compares the CRs for slow and fast compression for each channel of the simulated hearing aid, showing the arithmetic mean across subjects (and corresponding SDs), the geometric mean across subjects, and the maximum and the minimum values. For all channels above the first, the mean CR was greater for slow than for fast compression. The consequences of this are discussed later.

OPEN IN VIEWER

Table 1.

Comparison of Compression Ratios for SLOW and FAST Compression, for Each Channel of the Simulated Hearing Aid.

	SLOW					FAST
Channel	Ar. Mean	SD	GeoMean	Max	Min	Ar. Mean	SD	GeoMean	Max	Min
1	1.3	0.3	1.2	2.1	1.0	1.3	0.35	1.2	2.1	1.0
2	3.1	1.5	2.8	6.7	1.4	2.5	0.54	2.5	3.0	1.4
3	3.1	1.2	2.9	6.9	1.5	2.5	0.47	2.5	3.0	1.5
4	5.5	2.8	4.9	10	2.0	3.0	0.08	3.0	3.0	2.0
5	4.8	1.5	4.6	7.8	2.1	2.8	0.25	2.8	3.0	2.1

It should be noted that the slow compression would have had little effect on the ongoing fluctuations in amplitude for speech at a given level, but it would have resulted in the application of gains that varied with overall level, which is important because three overall levels were used (see later for details). Prior to the onset of the stimuli, the IG values in each channel were at their maximum values (equal to the IG at the CT). However, the use of an attack time of 50 ms for the slow compressor meant that the gain values were fairly rapidly reduced for the higher level stimuli.

All processing was performed offline, using at least 24-bit precision. Stimuli were generated via an Echo Indigo 24-bit sound card (Echo Digital Audio, Santa Barbara, CA), using a sampling rate of 44.1 kHz, and were presented via Sennheiser HDA200 headphones (Wedemark, Germany). The response of the headphones was measured on an acoustic manikin (Burkhard & Sachs, 1975), and digital filtering was used to adjust this response so that it resembled the diffuse-field response of the ear, thus ensuring that the correct IGs were obtained in the ear canal. The prescribed IGs and the CRs could be set very accurately in the software. Also, the output of the headphones for frequencies up to 10 kHz is rather consistent across individual ears (Han & Poulsen, 1998). Thus, the prescriptions were probably implemented more accurately here than would be the case in a clinical setting with real hearing aids.

IGs were reduced for the inexperienced subjects according to the CAM2A recommendations. The reductions averaged 1 to 3 dB for frequencies up to 4 kHz and increased to about 4 dB at 6 kHz, 8 dB at 8 kHz, and 10 dB at 10 kHz; the gain reductions varied only slightly with input level.

Stimuli

The speech stimuli for the English-speaking subjects were digitally recorded segments of running speech (connected discourse) obtained from one male and one female talker of British English. The recordings were a subset of those described by Moore, Stone, Füllgrabe, Glasberg, and Puria (2008). Examples of the spectra of the speech stimuli are given in that article. The recordings were edited to remove pauses for breath and extraneous sounds, but natural-sounding gaps of between 100 and 300 ms were left. One 4.8-s segment of speech was selected for each talker. The speech stimuli for the Polish-speaking subjects were digitally recorded meaningful sentences describing realistic situations. Four short sentences were concatenated to produce a 7.8-s sample of male speech and a 8.9-s sample of female speech. The level of the speech at the input to the simulated hearing aid was 50, 65, or 80 dB SPL.

The music signals were as follows: a 7.3-s segment of a jazz trio (piano, bass, and drums); a 5.6-s segment of a classical orchestra (including brass instruments and cymbals); a 3.5-s segment extracted from track 27 of the compact disc produced by Bang & Olufsen called “Music for Archimedes” (CD B&O 101), consisting of a xylophone (anechoic recording); and an 8.3-s segment of a male singing (a counter-tenor accompanied by guitar and recorder). The long-term average spectra of the first three signals are shown in Moore et al. (2011). For all four music signals, the input level to the simulated hearing aid was 50, 65, or 80 dB SPL.

Paired-Comparison Procedure

The procedure was similar to that described by Moore and Sek (2013). On each trial, the same segment of sound was presented twice in succession, once processed with fast compression and once with slow. The time interval between the two segments was 1.25 s. The possible orders were used equally often, and the order was randomized across trials. Within a given pair of sounds, the only difference between the sounds was in the compression speed; the input level was always the same. The subject was asked to indicate which of the two was preferred and by how much, using a slider on the screen. The continuum was labeled “1 much better,” “1 moderately better,” “1 slightly better,” “equal,” “2 slightly better,” “2 moderately better,” and “2 much better.” The next pair was stimuli was presented 1 s after the subject had entered their response.

Within each block of trials, all six types of signals were presented (Classical music, Jazz, Male singing, Percussion, Female speech, and Male speech), and the input level was kept constant. The 12 pairs of sounds (6 signal types × 2 presentation orders [fast compression first or slow compression first]) were presented in random order. Two blocks of trials were used for each subject and each input level. The order of presentation of input levels across blocks was random.

For a given trial, if fast compression (FAST) was preferred, the slider position was coded as a negative number, and if slow compression (SLOW) was preferred, the slider position was coded as a positive number. For example, if the order on a given trial was SLOW first and FAST second, and the participant set the slider position midway between “2 slightly better” and “2 moderately better,” the score for that trial was assigned a value of −1.5. The overall score for each compression speed and stimulus type (e.g., classical music) was obtained by averaging all of the subscores obtained for that speed and stimulus type. A score of −3 would indicate a very strong and perfectly consistent preference for FAST, whereas a score of +3 would indicate a very strong and perfectly consistent preference for SLOW. A score of 0 would indicate no preference.

Measurement of DLFs

The method used to measure DLFs was similar to that described by Moore and Vinay (2009) and Moore and Ernst (2012). A two-interval, two-alternative forced-choice task was used. One interval contained four successive 2000-Hz tones. The other interval contained four successive tones whose frequency alternated between 2000 Hz and 2000 Hz + Δf. The subject had to choose the interval in which they heard a fluctuation in pitch. Each tone lasted 400 ms, including 20-ms raised-cosine ramps. The silent gap between the tones within an interval was 100 ms. The silent gap between intervals was 400 ms.

Each ear was tested separately. Initially, the absolute threshold at 2000 Hz was measured using the two-alternative forced-choice method implemented in the TFS1 software (Sek & Moore, 2012). Three estimates were obtained for each ear, and the final estimate was based on the average of the three. For measurement of the DLF, the level of each test tone was set 30 dB above the measured threshold, that is, to 30 dB sensation level (SL). No level roving was used. The value of Δf was varied adaptively to determine the DLF. For the first run for each subject, the starting value of Δf was 100 Hz. For subsequent runs, the starting value of Δf was set to four times the threshold estimate for the first run. Following two correct responses in a row, the value of Δf was decreased, while following one incorrect response, it was increased. This procedure tracks the 70.7% correct point on the psychometric function. The value of Δf was changed by a factor of 1.953 (1.25³) until one turnpoint had occurred, by a factor of 1.5625 (1.25²) until the second turnpoint had occurred, and by a factor of 1.25 thereafter. The procedure continued until eight turnpoints had occurred. The threshold was estimated as the geometric mean of the values of Δf at the last six turnpoints. Four estimates were obtained for each ear, and the final estimate was obtained as the geometric mean of the four.

Measurement of Interaural Phase Sensitivity—The TFS-LF (Low Frequency) test

The TFS-LF test (Hopkins & Moore, 2010b; Sek & Moore, 2012) uses a two-interval, two-alternative forced-choice task, with a similar temporal structure to that described for the measurement of the DLFs. The tone durations and silent gaps were as described for the DLFs. All of the tones in one interval were in phase at the two ears. In the other interval, the first and third tones were in phase at the two ears, but the second and fourth tones had an interaural phase difference (IPD) of ϕ in the TFS only (the envelopes of the tones were synchronous in the two ears). The subject was asked to indicate the interval in which the tones appeared to move within the head.

The test frequency was fixed at 500 Hz. Initially, the absolute threshold at 500 Hz was measured for each ear using the two-alternative forced-choice method implemented in the TFS software (Sek & Moore, 2012). Three estimates were obtained for each ear, and the final estimate was based on the average of the three. The level of the test tones in each ear was set to 30 dB SL. The value of ϕ was varied adaptively to determine the IPD at threshold, using the same procedure as for the DLFs. The starting value of ϕ was 180°. The threshold was estimated as the geometric mean of the values of ϕ at the last six turnpoints. Five estimates were obtained, and the final estimate was obtained as the geometric mean of the five.

The adaptive procedure terminated if the procedure called for the largest possible value of ϕ (180°) twice before the second turnpoint or any time after the second turnpoint. This happened consistently for six of the subjects. The thresholds for those subjects were thus indeterminate but are plotted at an arbitrary value of 180° when the results are presented. When the adaptive procedure terminated, the procedure switched to a nonadaptive procedure where 40 trials were run with ϕ fixed at 180°. The obtained scores were all only a little above the chance value of 50% (range 53% to 58.5% correct).

Measurement of Interaural Phase Sensitivity—The TFS-Adaptive Frequency Test

To obtain an estimate of interaural phase sensitivity for all subjects, we used a modified version of the TFS-LF test in which the IPD was fixed at 180°, and the frequency of the test tone was adaptively varied to determine the highest frequency at which the task could be performed. The task is based on the assumption that, for each subject, there is a frequency above which sensitivity to interaural phase worsens markedly (Brughera, Dunai, & Hartmann, 2013; Hughes, 1940). The procedure is intended to track the highest frequency at which changes in IPD can be detected. The time pattern of the stimuli was the same as for the TFS-LF test. All subjects could perform the task when the frequency was made sufficiently low. We refer to the modified test as the TFS-AF test, where AF stands for adaptive frequency (Füllgrabe, Harland, Sek, & Moore, 2016). A comparable task has been used by Ross, Fujioka, Tremblay, and Picton (2007) and by Grose and Mamo (2010).

The starting frequency was 500 Hz. The level of the tones in each ear for this frequency was set to 30 dB SL, using the absolute threshold measured with the TFS software. Because the absolute thresholds varied with frequency, the absolute thresholds at other frequencies were estimated from the measured audiometric thresholds for each ear. For each audiometric frequency, the absolute threshold in dB HL was converted to level in dB SPL by adding the minimum audible pressure for monaural listening, as predicted by the loudness model of Moore and Glasberg (2007). Interpolation was used to estimate the absolute threshold in dB SPL at frequencies between the audiometric frequencies, and the level of the test tones in each ear was adjusted so that the SL remained constant at about 30 dB SL as the frequency was changed. The frequency was adjusted using the same rules as used for changing Δf in the DLF task, and the threshold was estimated as the geometric mean frequency at the last six turnpoints. Five estimates were obtained, and the final estimate was obtained as the geometric mean of the five. The lowest frequency allowed in the software was 30 Hz, and all subjects achieved thresholds of 200 Hz or above.

Compression Speed Preferences for Speech

As described earlier, a score of −3 would indicate a very strong and perfectly consistent preference for FAST, whereas a score of +3 would indicate a very strong and perfectly consistent preference for SLOW. A score of 0 would indicate no preference. The pattern of results was similar across levels. To get an overview of the individual preference scores, the scores were averaged across the three levels. The average preference scores for the male talker and the female talker were highly correlated (r = .91, p < .001). This indicates that the subjects were consistent in their ratings across talkers. In what follows, only the mean ratings across talkers are considered.

Figure 2 shows individual and mean preferences for speech. On average, SLOW was preferred over FAST, but only by 0.46 scale units. There were distinct individual differences. Ten subjects showed a preference for SLOW of 0.5 scale units or more, while five subjects showed a preference for FAST of 0.5 scale units or more. OPEN IN VIEWER

Compression Speed Preferences for Music

Again, the pattern of the preference scores was similar across levels, and the results were initially averaged across the three levels. The preference scores were reasonably consistent across all music types except for the solo percussion instrument. The scores for the latter were not significantly correlated with scores for any of the other music types. Also, the scores for the percussion instrument fell within a narrow range (−0.2 to +0.53) for all except two subjects. This may have happened because the percussion instrument was generally described as sounding unpleasant regardless of whether fast or slow compression was used. Therefore, in what follows, we consider only the mean scores across the three other music types: jazz, classical, and man singing with guitar.

Figure 3 shows individual and mean preferences for music. On average, SLOW was preferred over FAST, by 0.55 scale units. There were again distinct individual differences. Nine subjects showed a preference for SLOW of 0.5 scale units or more, and the rest showed very small preferences for SLOW or no clear preference. OPEN IN VIEWER

Similarity of Preferences for Speech and Music

Although the mean preference for SLOW relative to FAST was slightly greater for the music than for the speech stimuli, the pattern of preferences across subjects was highly correlated for the two stimulus types (r = .89, p < .01), as can be seen in Figure 4. However, it is noteworthy that subjects whose preference scores for speech were negative (indicating slight preferences for FAST) gave preference scores for music that were in the range 0 to 0.4, indicating no clear preference for SLOW versus FAST. Subjects with preference scores greater than 1 for speech, indicating relatively strong preferences for SLOW, also gave reasonably strong positive preference scores for music. OPEN IN VIEWER

Effects of Level on Preferences

Figure 5 shows mean preference ratings across subjects for each level, separately for the speech and music stimuli. An overall slight preference for SLOW occurred for all three levels for both stimulus types. A within-subjects analysis of variance was conducted on the preference scores with factors level and stimulus type. Neither of the main effects was significant, but there was a significant interaction between level and stimulus type: F(2,42) = 15.3, p < .001. The interaction reflects the fact that the relative preference for SLOW was greater for speech than for music at the 65-dB SPL level, but not at the two other levels. OPEN IN VIEWER

The pattern of results differs somewhat from that found by Moore et al. (2011) using a similar procedure, simulated hearing aid, and stimuli. They also found slight mean preferences for slow compression for both speech and music, but the effect occurred mainly for the input level of 80 dB SPL, and not for input levels of 50 and 65 dB SPL. The discrepancy may be related to the fact that their subjects generally had smaller hearing losses than those of the subjects used here. For example, at 4 kHz, the hearing losses of their subjects ranged from 20 to 65 dB, with a mean of 44 dB, whereas the hearing losses of the subjects used here ranged from 48 to 74 dB, with a mean of 59 dB. CAM2A prescribes higher CRs for greater hearing losses, so the effects of compression speed might have been larger and more noticeable for the subjects used here.

TFS-LF and TFS-AF Scores

Figure 6 is a scatterplot of scores for the TFS-AF test (in hertz) against scores for the TFS-LF test (in degrees). Scores for the latter are plotted as diamonds at 180° for the six subjects for whom the threshold could not be reliably determined because the adaptive procedure called for a value of ϕ greater than 180° (three of the diamonds are superimposed, as, by coincidence, three subjects had thresholds on the TFS-AF test of 200 Hz). The six subjects for whom thresholds could not be determined reliably using the TFS-LF test all had thresholds on the TFS-AF test close to 500 Hz or at 200 Hz, so it is not surprising that the adaptive procedure could not be completed using the TFS-LF test with a fixed frequency of 500 Hz. Of the remaining subjects, six had thresholds below (i.e., worse than) 500 Hz but were still able to complete the adaptive procedure in the TFS-LF test. For the 16 subjects who were able to complete both the TFS-LF and the TFS-AF tasks, there was a strong negative correlation between the two (r = −.91, p < .01), indicating good consistency across the two tests; good interaural phase sensitivity is associated with a low threshold in degrees on the TFS-LF test and a high threshold in hertz on the TFS-AF test. OPEN IN VIEWER

Previous results have suggested that sensitivity to interaural phase declines with increasing age (Grose & Mamo, 2010; Moore, Glasberg, Stoev, Füllgrabe, & Hopkins, 2012; Moore, Vickers, et al., 2012; Ross et al., 2007), even when audiometric thresholds remain within the normal range. For the present results, the TFS-AF thresholds were negatively correlated with age, as expected (r = −.2), but the correlation was not significant (p > .05), perhaps because the effects of age were swamped by the effects of hearing loss. The TFS-AF thresholds were significantly negatively correlated with the absolute threshold for the test ear at 500 Hz (r = .43, p < .05).

Relationship of Compression Speed Preferences to Sensitivity to TFS

Because we were testing the hypothesis that the relative preference for SLOW would increase with decreasing sensitivity to TFS, one-tailed tests were used to assess the significance of correlations. The logarithms of the DLFs for the test ears were weakly correlated with preference scores: r = .36, p = .05 for speech and r = .39, p < .05 for music. This is illustrated in Figures 7 and 8. Large DLFs, indicating poor sensitivity to TFS, were associated with greater preferences for SLOW. Scores on the TFS-AF test, which could be completed by all subjects, were not significantly correlated with compression speed preferences for either speech or music (both r = −.06, p > .05). The DLFs for the test ears were not significantly correlated with the TFS-AF test scores (r = −.17, p > .05). This is not surprising, as the two tests were conducted using stimuli in very different frequency regions. OPEN IN VIEWER

OPEN IN VIEWER

Figure 8.

Scatterplot of preference scores for music against DLFs at 2000 Hz.

The DLFs ranged from about 15 to 50 Hz, with a geometric mean of 24 Hz. Moore and Ernst (2012) measured DLFs for normal-hearing subjects using a similar task, except that they roved the level of every tone, whereas no roving was used here. Their DLFs at 2000 Hz for a level of 70 dB SPL ranged from 12 to 31 Hz with a geometric mean of 15 Hz. Their DLFs were, on average, smaller than those found here, despite their use of level roving, which would tend to increase the DLFs, consistent with the idea that the hearing loss of our subjects was associated with larger-than-normal DLFs. Some previous results have suggested that DLFs increase with increasing age (Konig, 1957; Moore, 2014). However, in the present results, there was no significant correlation between the DLFs and age (r = .08, p > .05), perhaps because the effects of age were swamped by the effects of hearing loss. The correlation between the DLFs and the audiometric thresholds at 2 kHz in the test ear was significant (r = .56, p < .01).

Comparison of CRs for SLOW and FAST

One complication in interpreting the results is that the CR was allowed to have any value up to 10 for SLOW, but the CR was limited to 3 for FAST. For all channels above the first, the mean CR was greater for SLOW than for FAST, as shown in Table 1. This might have affected the preference judgments. However, the slow compression had a release time that was so long that the compression would have had little effect on the ongoing amplitude fluctuations in the stimuli, even when the CR was above 3 (Stone & Moore, 1992, 2007). Therefore, the dynamic changes in the stimuli were effectively independent of the CR used for SLOW.

As noted earlier, the main effect of the slow compression was to compensate for differences in overall level of the stimuli. Because of the different maximum CRs used for SLOW and FAST and the way that CAM2A compensates for limitations in the CR, the output level was expected to be slightly higher for SLOW than for FAST for the 50 dB SPL input level and slightly lower for SLOW than for FAST for the 80 dB SPL input level; the output levels should have been well matched for SLOW and FAST for the input level of 65 dB SPL. To quantify these effects, the CAM2A software was used to generate a recommended fitting for an audiogram corresponding to the average for the test ears of the present subjects (with the CR restricted to 3 for FAST and 10 for SLOW). The simulated hearing aid was programmed with the two fittings. The output was then determined for each stimulus (except the percussion instrument) for each input level. The results are given in Table 2. The differences in output level between SLOW and FAST were in the expected direction but were generally small. Output levels were very well matched for the input level of 65 dB SPL.

OPEN IN VIEWER

Table 2.

Differences in Root-Mean-Square Output Level (FAST−SLOW) in dB for Each Signal (Excluding the Percussion Instrument) and Each Input Level.

	Input level, dB SPL
Stimulus	50	65	80
Male talker	−2.0	−0.1	4.3
Female talker	−2.0	−0.5	4.0
Classical	−2.0	0.8	6.4
Jazz	−1.9	−0.1	2.4
Man singing	−1.6	0.0	0.8
M	−1.9	0.0	3.6

If the preferences were determined by level (or loudness), one might have expected the preferences to differ most from zero for the lowest and highest levels, and to be least different from zero for the medium input level, for which the output levels were well matched. In fact, the overall mean preferences (across speech and music) showed a nonsignificant trend in the opposite direction; mean preference values were 0.41, 0.66, and 0.51 for the input levels of 50, 65, and 80 dB SPL, respectively (see Figure 5). It can be concluded that the preference judgments were probably not determined by level or loudness. The dominant factor was probably related to the fact that FAST reduced the ongoing envelope fluctuations in the stimuli, while SLOW did not.