Stimuli

Speaking American English at conversational level, a 23-years-old female produced tokens of the consonant-vowel-consonant (CvC) words /pæt/, /pæk/, and /kæt/, that were digitized (20 kHz) and saved as audio files with overall duration of 574, 565, and 540 ms, respectively (mean = 560 ms). On each CvC sound file (Fig 1), four time segments were delimited. The first is the release burst and aspiration noise of the pre-vocalic stop consonant, termed “Cv” burst, lasting 65, 75, and 75 ms for /pæt/, /pæk/, and /kæt/, respectively (mean = 72 ms). The second is the vowel, “v,” lasting 152, 165, and 187 ms, respectively (mean = 168 ms), and the third is the silence preceding the final stop (mean = 190 ms). The fourth is the release burst and aspiration noise of the post-vocalic stop consonant, termed “vC” burst, lasting 167, 135, and 88 ms (mean = 130 ms). In duration, the pre-vocalic (Cv) bursts are shorter than the post-vocalic ones (vC) but, to retain real-life conditions, the natural acoustic dimensions were not altered (except for the level increments). In what follows, the noise bursts are referred to as Cv or vC, and not as phonemes (e.g., /p/) because the burst is only a fragment of this phoneme.

Hearing-aid compression alters the acoustic properties of stop-consonant noise bursts in more ways than just the level [4, 5]. Thus, in this study, level increments confined to a single consonant noise burst (Cv or vC) were implemented off-line with a digital waveform editor that increased the burst level without changing its time waveform, spectrum, or any other CvC segment [5, 21, 22]. Achieving this with a real-time programmable attenuator is impractical because real-time attenuation changes from one to another segment of a CvC word (e.g., increasing only the initial burst level) produces audible clicks and interruptions. With the waveform editor, the amplitude, no (t), of the burst sample points was multiplied by (1+c), 0 ≤ c ≤ 5.1 to produce linear increments in RMS amplitude (as well as in pressure in dyne/cm²) much like those of linear amplification. For example, multiplication by a factor of 2 (i.e., c = 1.0), doubles the peak-to-peak amplitude of every zero-crossing undulation in the noise waveform, resulting in a 6.0-dB increment; a factor of 1.12 produces 1.0-dB level increments. All level increments were positive to approximate the effects of compression and consonant-vowel ratio enhancement on the bursts level. In general, the RMS-level difference, ∆L in dB, between a burst without and one with positive RMS-level increments is equal to 20 log (1+c); thus, in 5–11 steps of ≈0.4–3.0 dB each, ∆L ranged from 0.21 to 14.0 dB (these are the nominal dB values computed with the equation; the acoustic precision was ≈0.5 dB). With each CvC word, a 5–11 token series was created in which the tokens differed from one another only in the size of ∆L in the Cv burst, being 0 dB (i.e., c = 0) in the standard CvC and 0.21–14.0 dB in the comparison CvC. This series and a similar one with ∆L in the vC burst were termed “in-context” because the bursts were embedded in the respective CvC word. With the words /pæk/ and /pæt/, two additional burst-in-isolation series presented only the Cv or only the vC burst, excerpted from the CvC, but the waveforms and ∆L were identical to those of the in-context series. While excising the Cv burst from the vowel, the CvC waveform was stretched along the time axis and the excision was at a sample point with zero amplitude; a new sample point was chosen if, upon play back, the burst produced audible distortions or clicks. Altogether, a total of 36 bursts-in-isolation and 52 burst-in-context sound files were used in the study.

Task and procedure

In the 2-observation interval, 2-alternative forced choice (2I-2AFC) task of experiment 1, burst-in-context trials presented a standard CvC word (/pæk/ or /pæt/) without ∆L and a comparison CvC with a ∆L in Cv or vC, each on a separate observation interval, their order (1^st or 2^nd) being random with equal probability. Inter-observation and inter-trial intervals were 0.5 s and 5.0 s, respectively. Participants pressed keys on a keyboard to report which interval included ∆L and, except for this, the two CvCs of each trial were identical; thus, the task required attending selectively to the Cv burst while attempting to ignore the remainder of the syllable, which is challenging because the CvC word is perceived as a compact whole. To recover psychometric functions relating the correct probability, P(C), to the size of ∆L, a constant-stimulus method presented randomly with equal probability each of the 9–11 tokens (/pæk/ or /pæt/) comprising the series with increasing ∆L in Cv. The same constant-stimulus 2I-2AFC method measured discrimination of ∆L in the vC burst in CvC context, and in the two burst-in-isolation conditions presenting only Cv or vC. In counterbalanced order, the four conditions (Cv or vC, each in isolation or in CvC context) were studied in 0.5–1.5 hr sessions comprising 4–9 blocks of 50–90 trials each, with 5–7 min rest in between; audiograms were collected in the 1^st session. On average, each of the four conditions included 720–878 trials, except for two participants that required 1260–1710 trials, and prior to collecting data a short practice period was allowed; thus the LDTs reported herein represent performance of well-trained participants. The Institutional Review Board University (University of Oklahoma Health Sciences Center) approved the testing protocol; written consent was obtained from all participants.

Equipment and stimulus level

From the output channel of the computer D/A converter, the electrical stimulus waveforms were low-pass filtered at 8.5 kHz (Precision Filter LP 66/88), amplified (Crown D-75), attenuated, and presented monaurally over TDH-49 earphones at an overall SPL of 79–82 dB(A). Without ∆L, the bursts overall SPL was 59–62 dB(A), or ≈20 dB lower than the vowel, on average; the pre-vocalic burst of /pæt/ was only 50 dB SPL. Since they were shorter than the 300–400 ms temporal-integration time [23], the bursts ended before the ear reached full response, which could increase the LDTs; thus, moderately high presentation levels were used to make the bursts clearly audible [4]. All SPLs were measured in a 6-cm³ coupler and the fast, A-weighted network of a sound-level meter (Quest 155). The stimulus presentation, timing, and response recording were controlled by a Macintosh computer and LabVIEW II software (National Instruments); all listening tests took place in a sound-treated room. In all conditions, participants entered responses on a computer keyboard and feedback on a computer monitor tallied correct or error responses.

Participants

Participants were students or staff members at the University of Oklahoma Health Sciences Center recruited by way of personal contact or flyers posted in bulletin boards. The study enrolled a total of 17 healthy participants (13 female), native speakers of English, with chronological ages ranging from 21–47 years (mean = 26), one of whom withdrew from the study before completing the experiments and the data were discarded. Among the 13 participants of experiment 1, nine had hearing thresholds levels within the normal range (≤15 dB) at the audiogram octave frequencies. The other 4 participants had mild hearing loss and, to offset its effect, the SPL of the CvC words was increased to achieve ≥ 20-dB sensation levels, keeping constant the 20-dB difference between the bursts and the vowel level. The participants in experiment 2 were four college students, 23–25 years old, with normal hearing and experience in the experimental task. For listening, an hourly fee was paid to each participant. Upon completion of the experiments, the data were summarized and analyzed (June 2016-March 2017).

Results

Although they were familiar with the CvC words, in burst-in-context conditions, the participants did not exhibit best differential sensitivity to ∆L right from the outset of the listening tests. Instead, most of them needed upwards of 200–300 trials to learn to attend selectively to the burst with ∆L and a few needed many more. Because they activate object-based attention, CvC words are perceived as tightly integrated wholes rather than as a three-phoneme sequence (i.e., noise burst, vowel, noise burst), and most participants needed practice in order to discern ∆L accurately.

The results described below represent performance at or near the asymptotic upper limit, and include raw data from each participant since comparable evidence is not available in the open literature. Fig 2 shows representative psychometric functions for each of 9 participants tested in the four conditions with ∆L in the Cv or vC burst embedded in /pæk/ or in isolation, excerpted from /pæk/. To estimate LDTs, maximum-likelihood probits were fit to the data relating correct probability, P(C), to ∆L in dB; each data point is based on 80 trials or more. In all participants and burst conditions, accuracy increases monotonically with ∆L. For the Cv burst in isolation (Fig 2), the slopes are steep and ∆Ls ranging from 0–4.0 dB bracket the chance-to-perfect performance range. In contrast, for the Cv burst in word context, the data show more variability, the slopes are less steep, and ∆L ≥ 6.5 dB is needed to reach P(C) = 1.0; clearly, for ∆L in the Cv burst, differential sensitivity in CvC context is much lower than in isolation. Compared to those of the Cv burst, the psychometric functions for the vC burst (Fig 2B) are steeper both in isolation and in context, and the differences between these two conditions are less obvious.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Psychometric functions relating discrimination accuracy to the burst level increment.

In Fig 2, the correct probability, P(C), as a function of ∆L in dB is shown for each of 9 participants (one per row) listening to ∆L in the pre-vocalic burst of /pæk/ presented in isolation or in /pæk/ context (1^st & 2^nd column on the left-hand side). Smooth curves are maximum-likelihood probits that come closest to the data points. In the same format, the 3^rd and 4^th column on the right side show functions for ∆L in the post-vocalic burst presented in isolation or embedded in /pæk/. See Supporting Information S1 Fig for further details.

https://doi.org/10.1371/journal.pone.0190956.g002

Fig 3 shows LDTs estimated from probits at P(C) = 0.75 and 0.85 (termed LDT_0.75 and LDT_0.85), to illustrate the consistency of the trends across two accuracy levels. The average LDT_0.75 was 3.24, 1.84, 2.47, and 1.93 dB for Cv in context, Cv in isolation, vC in context, and vC in isolation; respectively, the average LDT_0.85 was 4.74, 2.95, 3.71, and 2.79 dB. Repeated measures ANOVA, with burst temporal position (initial Cv versus final vC) and presentation format (in-context versus isolation) as factors revealed significant main effects of both factors and a significant interaction. With the LDT_0.75, significant effects obtained with temporal position, F(1, 8) = 9.38, MSE = 1.036, p = .016, context-isolation, F(1, 8) = 19.45, MSE = 8.341, p = .002, and the position-context interaction, F(1, 8) = 8.17, MSE = 1.668, p = .02. With the LDT_0.85, the position, F(1, 8) = 7.02, MSE = 3.148, p = .029, context, F(1,8) = 21.32, MSE = 16.52, p = .002, and their interaction, F(1, 8) = 5.81, MSE = 1.747, p = .042, also achieved significance. The trends are much the same at the two accuracy levels. Pairwise comparisons (2-tail Student’s t-test) of LDT_0.75 showed significant differences between Cv context and Cv isolation (p = .001), Cv context and vC context (p = .01), and Cv context and vC isolation (p = .00); that is, the context, temporal-position, and interaction effects remain significant if the confidence level takes into account multiple comparisons (p = .017). In spite of the duration difference between the 75-ms Cv and the 135-ms vC burst, in isolation, the LDT_0.75 did not differ significantly (p = .52); that is, the shorter duration of the Cv burst cannot account for the LDT differences observed in burst-in-context conditions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Level Discrimination Thresholds (LDTs).

The LDTs (Y axis) measured in experiment 1 with the words /pæk/ and /pæt/ are depicted in the top (A) and middle panels (B), respectively; the bottom panels (C) depict the LDTs measured in experiment 2. Panels on the left- and right-hand side depict LDTs estimated at the 0.75 and 0.85 correct probability, respectively. In the top panels, for level increments in the pre-vocalic (Cv) burst of /pæk/, bars labeled PC and PI depict average LDTs (ordinate) for Cv in context and Cv in isolation; for level increments in the post-vocalic (vC) burst of /pæk/, bars labeled KC and KI depict the LDTs for vC in context and vC in isolation. For ease of comparison, on the left-hand side panels, the dashed horizontal lines depict the LDT for level increments encompassing the entire CvC word measured by [8]; since this LDT was estimated only at a correct probability of 0.71, it is not included in the right-hand side panels. In all panels, error bars depict 1.0 standard error below the mean. Middle panels (B): for level increments in the pre-vocalic (Cv) burst of the word /pæt/, bars labeled PC and PI depict average LDTs (Y axis) for Cv in context and Cv in isolation; for level increments in the post-vocalic (vC) burst, bars labeled TC and TI depict the corresponding thresholds for vC in context and vC in isolation. Bars labeled PC+T and TC+P depict LDTs measured in the word-context condition that interleaved level increments in the pre- (Cv) or post-vocalic (vC) burst within each trial block (temporal-position variability). Bottom panels (C): for level increments in the pre-vocalic (Cv) burst of /kæt/, /pæt/, or /pæk/, bars labeled CAT-K, PAT-P and PACK-P depict average LDTs (Y axis) for Cv in context; bars labeled CAT-T, PAT-T, and PACK-K depict the corresponding LDTs for level increments in the post-vocalic burst (vC). Since all six conditions were interleaved within each trial block, the task included variability in the temporal position of the level increment and in the consonant place of articulation. See Supporting Information S2 Fig for further details.

https://doi.org/10.1371/journal.pone.0190956.g003

To determine if the above trends generalize to other words, in a new participant crew (n = 4), differential sensitivity to ∆L in the stop-consonant noise bursts of /pæt/ was studied also. With the pre-vocalic burst in context, ∆L discrimination was far more difficult with /pæt/ than with /pæk/, and this enabled measuring training effects since it required many more discrimination trials. A repeated-measures ANOVA did reveal significant P(C) differences, F (1, 5) = 6.84, MSE = 1629.76, p = .047, between the initial and final set of 200 trials, each presenting a standard and a comparison /pæt/; that is, mastering the task of discerning ∆L in this condition did require significant training, even though the bursts were embedded in familiar words. Measured as with /pæk/, Fig 3 (middle panels) shows average LDTs for ∆L in the Cv or vC burst each presented in context or isolation, plus an additional subsidiary condition described below. The average LDT_0.75 was 9.72, 2.67, 2.37, and 1.99 dB for Cv in context, Cv in isolation, vC in context, and vC in isolation; the respective average LDT_0.85 was 12.27, 3.95, 3.63, and 2.87 dB. Repeated measures ANOVA, with burst temporal position (initial Cv versus final vC) and presentation format (in-context versus isolation) as factors, revealed significant main effects of both factors and a significant interaction. For the LDT_0.75, significant effects obtained for temporal position, F(1, 3) = 35.13, MSE = 64.56, p = .01, context-isolation, F(1, 3) = 34.42, MSE = 55.38, p = .01, and their interaction, F(1, 3) = 18.40, MSE = 44.55, p = .023. Repeated measures ANOVA of the LDT_0.85, showed the same significant effects: temporal position, F(1, 3) = 27.97, MSE = 94.53, p = .01, context-isolation, F(1, 3) = 64.49, MSE = 82.22, p = .04, and their interaction, F(1, 3) = 15.36, MSE = 57.19, p = .03. Pairwise comparisons (2-tail Student’s t-test) of LDT_0.75 showed significant differences between Cv context and Cv isolation (p = .003), between Cv context and vC context (p = .014), and between Cv context and vC in isolation (p = .01). That is, the context, temporal-position, and interaction effects remain significant if the confidence level takes into account multiple comparisons (p = .017). The LDT_0.75 difference between vC context and vC isolation was not significant (p = .16) because ∆L discrimination was very precise in both conditions. For bursts in-isolation, the difference between the average LDT_0.75 of the 167-ms vC burst and the 65-ms Cv burst was not significant (p = .12)

Following the above, the same participants were run in a subsidiary condition that interleaved, within each trial block, the two burst-in-context conditions of /pæt/ and, at random but with equal probability, ∆L was either in the Cv or vC burst, but not in both. On a given trial, the observation interval presenting the burst with ∆L was also chosen randomly with equal probability. The goal was to determine if variability in the temporal position of ∆L would increase the LDTs since it required switching attention between the Cv and vC burst, which could hamper the ability to separate the burst auditory grouping from that of the vowel. In the subsidiary condition, testing required 8–12 blocks of 96–120 trials each, spanning 1–2 test sessions. Since all participants had practice in the preceding four conditions, this pre-training could generalize to the present condition.

With variability in the temporal position of ∆L, the average LDT_0.75 was 10.82 and 2.61 dB for the Cv and vC burst; respectively, the average LDT_0.85 was 13.25 and 3.83 dB (Fig 3, middle panels). A repeated measures ANOVA contrasting the LDT_0.75 of interleaved against the non-interleaved burst-in-context condition, again revealed a significant temporal-position effect, F(1, 3) = 35.63, MSE = 201.94, p = .01, but neither the interleaving effect, F(1, 3) = 2.06, MSE = 7.20, p = .25, nor their interaction, F(1, 3) = 7.80, MSE = 4.86, p = .07, were significant. Variability in the temporal position of ∆L and switching attention across the two bursts did produce further LDT elevations but the small effects did not reach significance. This outcome is not that unexpected because this variability does not alter the perceptual grouping of the noise bursts or of the vowel, and the extensive training in the preceding, non-interleaved condition could generalize to the interleaved condition. Lack of significance could be due also to the small sample sizes, however our main goal was showing that with variability in the temporal-position ∆L participants can perform much like in no-variability conditions.

For non-speech broadband noise, 60-63-dB overall SPL, average LDT_0.71 of 2.33 and 0.9 dB have been reported [24, 25] for 30-ms and 500-ms bursts, a duration range that encompasses the 65–187 ms of the pre- and post-vocalic (Cv and vC) consonant noise bursts used here; for ∆L in the whole CvC, the average LDT_0.71 is 2.2 dB [8]. Compared to the sensitivity to ∆L in like-duration random noise, the average LDT_0.75 for Cv and vC noise bursts in isolation (i.e., 1.84 & 1.93 for /pᴂk/, and 2.67 & 1.99 for /pᴂt/) fall within the LDT range for random noise and for the whole CvC; this in spite of performance-level, procedure, and stimulus differences between the previous and the present study. For both pre- and post-vocalic bursts in isolation, differential sensitivity to ∆L approaches the upper psycho-acoustic limit, implying that their duration difference could not account for the much larger temporal-position and context-effects reported here. With two CvC words (/pᴂk/ and /pᴂt/), the present results show significantly larger LDTs for bursts in word context than in isolation (a context-interference effect), for pre- than post-vocalic bursts (a temporal-position effect), and the largest LDTs for pre-vocalic bursts in context (an additive interaction of initial position and context). The temporal-position effect is much larger for /pᴂt/ than for /pᴂk/ probably because, in the former, the pre-vocalic burst (without ∆L) is ≈9.0-dB softer than the average; however, for this burst, only the in-context LDTs were enlarged, the in-isolation LDTs were not. Indexed in terms of LDTs, the perceptual significance of ∆L in noise bursts embedded in CvC words is much lower for the pre- than the post-vocalic burst; however, perceptual significance was not affected by variability in the temporal position of ∆L. The hypothesis that object-based attention modulates the discrimination of level increments in consonant noise bursts is supported by the superior discrimination accuracy observed with post-vocalic bursts, the auditory grouping of which would be more coherent than that of pre-vocalic bursts. Also consistent with object-based attention is the lower discrimination accuracy for bursts in context than in isolation, reflecting the inability to fully segregate the vowel from the noise-burst auditory grouping. Additional support accrues from the significant interaction between temporal-position and the context effect, and the similarity of the trends observed in conditions without or with variability in the temporal position of ∆L, since the latter does not alter the bursts auditory grouping.

Task, stimuli, and procedure

The main conditions of experiment 2 employed the same 2I-2AFC task of experiment 1, but within each trial block the standard and comparison were tokens of one of three CvC words (/pᴂt/, /pᴂk/, or /kᴂt/) sampled randomly but with nearly equal probability within each trial block. In addition, ∆L occurred randomly with equal probability in either the Cv or the vC burst of each word. Although larger than in experiment 1, the word set in experiment 2 remained small because the constant-stimulus method required a ∆L series of 8–9 tokens both for the Cv and the vC burst of each word, resulting in a total of 52 sound files. In addition, as many as 4600 trials were needed to obtain trial densities sufficient to compute reliable P(C) estimates with each sound file (or ∆L size). Data collection required 4–5 sessions, each consisting of 6–8 blocks of 90–110 trials.

Results

As in experiment 1, for each participant, the LDT_0.75 and LDT_0.85 were estimated from maximum-likelihood probits separately for the Cv and the vC burst of each word. For the Cv burst, the average LDT_0.75 was 9.98, 7.00, and 4.79 dB for /kᴂt/, /pᴂt/, and /pᴂk/; for the vC burst, the average LDT_0.75 was 3.54, 3.14, and 3.25 dB, respectively (Fig 3, bottom panels). A single-factor repeated-measures ANOVA comparing average LDT_0.75 for the Cv and vC burst showed a significant effect of temporal position for the three-word set, F(1, 2) = 11.47, MSE = 70.05, p = .01. A second repeated-measures ANOVA with word (3 levels) and temporal position (2 levels, Cv or vC) as factors revealed a significant effect of word, F(1, 2) = 13.02, MSE = 11.47, p = .02, but not of temporal position, F(1, 2) = 6.18, MSE = 70.05, p = .13; that is, for the latter factor, the LDT_0.75 differences were significant for the word set, but not so for individual words, because of the lower trial density. Based on the 95% confidence interval, the between-word LDT_0.75 differences were significant for the Cv but not for the vC burst; discerning ∆L in the Cv burst was far less accurate with /kᴂt/ than with /pᴂk/ [CIs = 6.94, 13.01], [3.51, 6.07]. The between-word LDT_0.85 differences were not significant.

As mentioned above, experiment 2 did not measure LDTs for bursts in isolation; thus, the isolation-versus-context effect was assessed against the isolation LDT_0.75 measured in experiment 1. In spite of the sample-size differences, an ANOVA comparing the average bursts-in-context LDT_0.75 of experiment 2 to the average burst-in-isolation LDT_0.75 of experiment 1 did reveal highly significant context-interference effects for both /pᴂk/, F(1, 22) = 21.50, MSE = 20.44, p = .00, and /pᴂt/, F(1, 12) = 9.80, MSE = 25.90, p = .01, because in experiment 2 the LDTs were 1.0–5.0 dB larger than those of experiment 1 (see Fig 1).

Subsidiary conditions

The above participants underwent testing in two subsidiary conditions. The first was a control presenting only /pᴂt/, with variability in the temporal position of ∆L (either Cv or vC) but not in the consonant that produced the burst. In the same participants, the difference between the LDT measured in this and in the preceding condition would estimate the effect of consonant variability alone. In the /pᴂt/ only condition, the average LDT_0.75 of 7.52 and 3.38 dB for the Cv and vC burst was not significantly different from the average LDT_0.75 of 7.00 and 3.14 dB measured in the three-word condition, F(1, 2) = 0.332, MSE = .432, p = .623. In addition, for the condition with /pᴂt/ only, the LDT_0.75 values were not significantly different from those measured in the temporal-position variability condition of experiment 1. A between-groups ANOVA showed a significant effect of temporal-position, F(1, 5) = 434.85, MSE = 507.32, p = .00, but not of group, F(1, 5) = 4.65, MSE = 5.43, p = .084; that is, neither the within- nor the between-group ANOVA revealed significant differences. In the aggregate, the results suggest that, for bursts in CvC context, differential sensitivity to ∆L does not decrease further by combining variability in the consonant with variability in the temporal position of ∆L; however, this conclusion is restricted to the word /pᴂt/ (the subsidiary condition did not include /pæk/, and /kæt/) and the small, single-vowel CvC set of the present condition. This outcome is within what could be expected because in order to discern ∆L, the perceptual grouping of the noise bursts and the vowel would be much the same across all three words. Further research is needed to corroborate this conclusion in a larger speech corpus and participant sample.

The second subsidiary condition assessed how burst-level discrimination depends on having the option to notice, within each 2I-2AFC trial, the contrast between the standard and the comparison bursts of the CvC word. Such within-trial contrast is not possible in a single observation-interval 2AFC task (termed SI-2AFC) in which trials present only the standard or the comparison CvC, in random order across trials; however, the perceptual segregation of the noise-burst from the vowel is still possible. The SI-2AFC task fosters attention to within-word level differences between the bursts, and reliance on memory traces of the CvC amplitude-modulation pattern (Fig 1). To discern ∆L in the noise bursts, participants can focus on the level relation between the Cv and the vC burst of each CvC; that is, on within-word level comparisons, rather than solely on standard-to-comparison burst-level differences across the observation intervals of the 2I-2AFC trial; alternatively, participants can rely on across trials burst-level comparisons. If discrimination of ∆L depends mainly on within-word comparisons and internalized CvC memory representations, allowing a single CvC observation per trial should not be much different than allowing two observations. If discrimination of ∆L were based mainly on noticing differences between the standard and comparison presented on each trial, omitting one or the other from each trial would degrade performance. To test these predictions, in the context of the word /pᴂt/, the LDT_0.75 for ∆L in the pre- or post-vocalic burst was measured also with a SI-2AFC task presenting only the standard or only the comparison on each trial, in random order from trial to trial; SI-2AFC discrimination accuracy was compared to that of the 2I-2AFC task of the previous subsidiary condition.

Context effect and perceptual grouping of noise bursts and vowel

For noise bursts embedded in CvC words, LDTs are significantly larger than those of bursts in isolation and of whole CvCs, meaning that the embedding significantly decreases the differential sensitivity, and this context interference is more pronounced for the pre- than the post-vocalic burst. The LDT enlargement observed in test conditions that maximize discrimination accuracy is likely to be greater in conditions approaching realistic speech-recognition conditions. To account for discrimination of ∆L in stop-consonant noise bursts of CvC words, the putative perceptual strategy would need to take into account the strong effects of context and of temporal position. In addition, such strategy would need to consider the weak effects of variability (in temporal position and consonant place of articulation) and of allowing only one observation interval that eliminates the benefit of contrasting the standard to the comparison CvC within each trial. From the stand-point of object-based attention [28, 29], the phoneme sound sequences (e.g., noise bursts, vowels) embedded in words (e.g., /pᴂt/) are perceived as a “unitary object” or “integrated whole” that makes it more difficult to attend selectively to individual phoneme sounds than to the full word; this is a top-down, cognitive effect that works against the breaking down of words into separate phonemes. In agreement with this view, for noise bursts in CvC context, the LDTs are significantly larger than those for whole CvCs. In addition, for bursts in CvC context, most participants needed upwards of 300 training trials (each including a standard and a comparison CvC, or 600 total) to master the discrimination task. That is, participants had to "learn" to attend selectively to the pre-vocalic noise bursts, though they were fully acquainted with the words, and could discern one from the other without errors. With consonant-vowel syllables, failure to attend selectively to specific phonetic segments has been reported previously [14]. A similar perceptual integration or grouping of the target and masker (i.e., burst and vowel) has been proposed to account for the elevation of level discrimination thresholds for pure tone targets under backward and forward masking [9, 10]. Increasing directly with the temporal proximity and similarity between the target and the masker [30], the perceptual-integration or grouping strength could account in part for the context effect; i.e., the enlarged LDTs of pre- and post-vocalic noise bursts embedded in CvC context.

Leakage of vowel-level information in the decision variable that encodes the noise-bursts level difference

The perceptual grouping of the burst and vowel can decrease differential sensitivity owing to a confounding of information. While deciding if a noise-burst includes ∆L, the task-irrelevant 20-dB jump from the Cv noise burst to the vowel commingles with the task-relevant information that encodes the small level difference (∆L) between the standard and comparison noise burst (across observation intervals). That is, the decision is based on a combination of task-relevant noise-burst level information and task-irrelevant vowel-level information [9, 10]. For the prevocalic noise burst, this combination seems almost inevitable because the observer is attempting to discern a small task-relevant ∆L in the noise burst while having to ignore the larger ∆L from the noise-burst to the vowel (see Fig 1), which is irrelevant to the task. In other words, in the present 2AFC tasks, discerning ∆L in pre-vocalic stop-consonant bursts depends partly on the within-observation interval (or within-word) level fluctuation (≈20 dB) from the Cv noise burst to the vowel. For instance, to be noticeable, the level increment in the target burst would have to increase the combined loudness of the pre-vocalic burst and the vowel, thereby decreasing the signal-to-noise ratio of the sensory information that encodes the level increment in the noise-burst target, because the vowel information is task irrelevant. Burst level discrimination would depend partly on the level difference between the noise bursts and the vowel rather than solely on the level difference between the standard noise burst presented in one observation interval and the comparison noise burst presented in the other interval [31, 32]. This limitation in the ability to keep apart the burst from the vowel is consistent with evidence showing that the level of prevocalic bursts is perceived in relation to the vowel level [1, 3]. With stop-consonant noise bursts, differential sensitivity to ∆L was much the same whether the 2AFC task included only one or two observation intervals; this finding suggest that participants can discern the burst ∆L partly on the basis of a within-observation interval comparison of the Cv and vC level. In addition, with familiar CvC words such as the ones used here, the memory representations can be highly accurate and function as remembered standards against which the CvC stimuli can be compared [16]; however, relying on these internalized standards does not nullify the context-interference or temporal-position effects.

Reorganizing the perceptual grouping of auditory events: Temporal-position and variability effects

As mentioned earlier, discerning ∆L in the noise bursts requires ignoring the linguistic information and attending selectively in order to reorganize the auditory events into three separate auditory objects: the to-be-attended pre-vocalic (Cv–Cv) or post-vocalic burst pair (vC–vC), and the to-be-ignored vowels (v–v). Discrimination accuracy would depend partly on how effectively the participant can set apart the noise-burst from the vowel object. The much larger LDT elevations observed with pre-vocalic than post-vocalic bursts stem from differences in the auditory streaming of the bursts and in the acoustic linkage between the bursts and the vowel (see Fig 1). Owing to the decay of vowel amplitude and the ≈190-ms silence that precedes the stop-closure release, the stream of post-vocalic bursts disassociates from the vowel to form a coherent auditory object that activates object-based attention, enabling precise ∆L discrimination. In contrast, owing to their close temporal proximity with the vowel, the stream of pre-vocalic bursts defines an ambiguous auditory object, which hinders object-based attention and ∆L discrimination. The temporal position effect provides strong support to the hypothesis that object-based attention modulates the discrimination of level increments in stop-consonant noise bursts.

In the context of the above, it is worth noting that the temporal-position effect is not merely a result of peripheral auditory masking; in other words, claiming that degraded differential ∆L sensitivity arises because the pre-vocalic burst is more susceptible to temporal masking and loudness contrast effects. With the present CvC words, it seems unlikely that the vowel could decrease the detectability of the pre-vocalic burst because this temporal backward masking is negligible when the probe approaches the 65–75 ms of pre-vocalic bursts [33, 34] presented well above detection threshold (59–62 dB SPL). Likewise, negligible decrease in detectability would be expected from forward masking of the post-vocalic burst by the vowel masker because, in addition to its long duration and high level, this burst is separated from the vowel by the ≈190-ms silence preceding the burst release. The inferior differential sensitivity for bursts in CvC context may not be explained in terms of low burst audibility stemming from temporal masking. Previous research [35] on backward and forward masking of 20-ms filtered noise bursts by synthetic-vowel maskers showed that peripheral temporal masking has little or no effect in the perception of natural stop-consonant bursts. The relatively wide bandwidth of the noise burst (1.5–7.0 kHz) ensures detectability in spite of being lower in spectrum level.

In agreement with the above perceptual strategy, variability in the temporal position of ∆L and in the burst place of articulation had minimal effects on differential sensitivity to ∆L in the noise bursts. Because it does not alter the perceptual grouping of the noise bursts, variability along one or both of these dimensions does not degrade the perceptual processes involved in this strategy. This includes the perceptual grouping of the noise-bursts and the vowels, extricating the bursts from the vowel, and the intrusion of vowel-level information in the decision process linked to the target ∆L. The time separation between the noise burst and the vowel is effective to modulate their integration, but between-trials variability in the temporal position ∆L is not. However, it is important to keep in mind that the variability in both the consonant place of articulation and in the temporal position of ∆L was quite limited: the stop-consonant set included only three CvCs all having the same vowel /ᴂ/, the ∆L could occur in only two temporal positions, and the participants had considerable task training. Making definite conclusions requires more comprehensive experiments having wider variability ranges along the above dimensions.

Implications for dynamic-range compression

A major (but not the only) effect of dynamic range compression is to increase the level of stop consonant noise bursts. For instance, for vC syllables consisting of the vowel /i/ and a post-vocalic stop-consonant, a ≈7.0-dB increase in the consonant-vowel intensity ratio (CVR) was observed [4] with compression release times of 12- and 100-ms, and a 2.0-dB increase with 800-ms release times. In the present study with CvC words consisting of the vowel /ᴂ/ and pre- and post-vocalic stop consonants, the average LDT_0.75 for ∆L in post-vocalic bursts ranged from 2.37–2.47 dB (see experiment 1). The present LDTs suggest that with 800-ms release time, the CVR increase would be perceptually insignificant, which is in agreement with the negligible intelligibility improvement at the 80-dB SPL presentation level. The LDTs are in agreement also with the significant intelligibility improvement observed with compression release times of 12 or 100 ms [4]. Taking into account the large methodological differences between the two studies (e.g., vowel /ᴂ/ versus /i/, pre- versus post-vocalic noise bursts, discrimination versus intelligibility performance indices), the agreement is noteworthy, suggesting that differential sensitivity for acoustic distortions stemming from compression could predict perceptual significance, and perhaps intelligibility improvement [6, 8]. However, the present differential-sensitivity estimates suggest also that this agreement is by no means general and highlight the need to consider contextual variables such as whether the stop consonant is pre- or post-vocalic. For example, even if the compression acoustic distortions were the same for pre- and post-vocalic stop-consonant noise bursts (which are not), the distortions perceptual significance would not be equal because their significance depends on temporal position. For ∆L in stop-consonant noise bursts an important contextual variable is whether the burst is followed immediately by a large increase in the vowel level (pre-vocalic) or preceded by a large drop in vowel level and silence (post-vocalic); the perceptual significance of such level increments is much lower for pre- than for post-vocalic stop consonants. Acoustic indices such as CVR or EDI do not consider the effects of these variables. Differential sensitivity is not the same across the pre- and post-vocalic consonant, not across stop consonants differing in place of articulation (bilabial versus velar).

The present results highlight the need to take into account the perceptual significance of level differences associated with dynamic-range compression or CVR enhancement. In CvC words, the perceptual CVR of pre-vocalic consonant noise bursts is much lower than the acoustic CVR; for the post-vocalic consonant noise burst, the perceptual and acoustic CVR are similar. Efforts to improve the processing of noise bursts should focus more on the pre-vocalic than in the post-vocalic stop consonants. In regard to compression, the most significant implication of the present results is in showing that differential sensitivity to ∆L in stop-consonant noise bursts of CvC words depends strongly on cognitive processing, including perceptual grouping, selective attention, and memory. This sensitivity does not depend solely on acoustic variables such as CVR or EDI or on peripheral auditory processing; deficits in the perceptual processing of these noise bursts could result partly from failure to attend rather to hear them.

Limitations of the present study

Since this could well be the first attempt to measure discrimination thresholds for level increments in stop-consonant noise bursts, the main goal was to determine how general and robust are the main effects (i.e., temporal-position and context); thus, a relatively wide range of conditions was examined in relatively small participant samples. Also innovative is the effort to account for the results in terms of perceptual grouping and object-based attention. Overall, the outcome of the study is quite promising, but there are important limitations the most prominent of which are the relatively small samples sizes, and the small speech corpus used (CvCs only). Future studies are needed recruiting much larger samples and employing a more representative speech corpus. Exploring the possibility of using a psychophysical method that is more efficient than the constant stimuli is also called for; because this method is far too demanding in terms trials needed to obtain reliable estimates of discrimination accuracy, testing large participant samples is almost prohibitive.