Cueing in a perceptual task causes long-lasting interference that generalizes across context to affect only late perceptual learning and is remediated by the passage of time

Perceptual learning, the improvement in sensory discriminations with practise, is also subject to stimulusspecific interference from temporal jitter in a learning session or manipulations applied between or immediately after sessions. We demonstrate a novel form of perceptual interference where even a brief cueing exposure to a complex speech-in-noise task produces a forward interference on subsequent speechin-noise learning. This potent interference generalizes across cueing context but specifically affects only late learning in the subsequent task, is resistant to the remediating effects of sleep and persists across an overnight delay involving sleep, and can be evoked by a single exposure 1 day before the learning. Learning in the speech-in-noise task is due to generalized improvements in discriminating and extracting signals (speech) from noise and we hypothesize that the forward interference represents interference with improvements in access to higher-level representations in rapid perception of ecologically-familiar complex signals such as speech from background noise. Perceptual learning underlies improvements in sensory discrimination abilities with training and skilled sensory detection, as demonstrated by a meteorologist identifying weather patterns from synoptic charts, air traffic controllers being able to monitor flight patterns of many aircraft simultaneously on radar or other display screens, or an oenologists expertise in differentiating wine flavours. In vision this can be demonstrated with practise-induced improvements in orientation, texture and vernier discrimination tasks, or odd-one-out tasks. In audition humans improve with practise on tasks ranging from the discrimination of simple sounds differentiated by frequency, timing or interval 1-11 to discrimination of relatively complex sounds like phonemes and words . These skills underlie speech identification and similar practiceinduced improvements may be expected in complex speech tasks. In keeping with this it had been demonstrated that practise with a specific talker’s voice helped in subsequent recognition of isolated words and words-in-sentences in that voice , and recently we 19, 20 have shown rapid and substantial learning in identification of speech in noise, using natural conversational speech in the presence of different background noises, consistent with a previous isolated observation of improvement in the first two sessions of a German sentences-in-noise test . Different sentences were used in all learning sessions in our study and hence the learning reflected a generalized improvement in the ability to extract complex, natural signals from competing sounds. Perceptual learning is also subject to interference. This can occur by introducing temporal jitter in the presentation of stimuli within a training session that involves learning of multiple stimuli along the same perceptual dimension . Between-sessions interference can occur in multi-stimuli learning if temporally-jittered stimuli are introduced within 4 h of training with non-jittered stimuli . In singlestimulus learning, between-sessions interference can be elicited by training with a similar hyperacuity task with a small variant, within 1 h of a previously-trained hyperacuity task 23 or by repeated within-day training , and this appears to occur through effects in primary sensory cortex 27, . All these cases involve interference with the consolidation of perceptual learning, a process that often involves sleep (e.g., 29, 25), and it is therefore not surprising that even a short nap can reverse perceptual deterioration occurring with repeated within-day training . Perceptual learning consists of an early, rapid phase of improvement and a later slowlydeveloping phase 1, 3, 4, 6-8, 10, 11, 13, 15, 16, 25, 26, 30-36 (see review by 37). The former is generally assumed likely to reflect procedural learning, the initial learning that occurs when acquiring “how to” aspects of a new skill or the response demands of a task, and the latter to reflect perceptual learning, the improvement reflecting changes in sensory function at a cortical level. More recent evidence indicates that even the early phase consists of substantial perceptual learning 6 and involves rapid changes in primary sensory cortex 3, 33, 36, 38-40 which are attention-dependent and only preserved with further practice . This early phase differs from the late phase in that it can generalize to non-trained stimuli whereas the late phase is stimulus-specific 6, 8, 41 and it has been argued that this makes the early phase the more important for applied (clinical) situations . The differences between these two phases of perceptual learning are still to be delineated and, in particular, it is not known how interference affects them; in fact interference is often taken to be a third and separate process . In the present study we demonstrate a novel form of interference where very brief cueing with a complex speech in noise task affected subsequent learning in that same general task with different stimuli. The suppression was long-lasting, could be triggered by single cueing session, and generalized across background context of the cueing task but then selectively affected only late perceptual learning. RESULTS Cueing results in context-specific improvement in baseline in the learning task and context-general interference of only late learning In the first experiment the effect of Cueing in the speech-in-noise task (Cueing with two different sentence lists separated by 2 mins, in one type of noise masker) was examined 30 mins later on the Learning speech-in-noise task (with six different sentence lists also spaced 2 mins apart). The different noise backgrounds (Babble Noise, BN, n = 18; Speech Weighted Noise, SWN, n = 23; modulated SWN, SWN-BNenv;, n = 19; see Methods) in the Cueing condition affected the level of difficulty in identifying sentences. Figure 1a plots Speech Reception Thresholds (SRTs; the signal-to-noise ratio for 50% discrimination and identification) for the first Cue list in the three cued groups (which was the same list of sentences in all groups albeit with sentences presented in random order between participants across groups); for comparison data for the first list in the Uncued group (n = 21) (namely, the first Learning-task list, in background BN) is also shown. Statistical analysis confirmed that the different noise masker caused different levels of difficulty in identifying speech-in-noise. The two BN groups were pooled since their data should be equivalent: all sentence lists in this study were selected (see Methods) to have similar discrimination and identification characteristics and this was confirmed here by comparison of SRTs for the first Cue list in the group trained in BN against SRTs for the first [Learning] list in the Uncued group (Students t-test; t = -2.0, df = 37, p>0.05). Then comparing data for the three noise types, there was a statistically significant difference between SRTs (Fig. 1a; 1-way ANOVA: F2, 78 = 41.9, p < 0.0001): the multi-talker BN background was the hardest condition (most positive SRTs); the easiest condition (most negative SRTs) was the background of modulated Speech Weighted Noise (SWN-BNenv) although with the greatest scatter; and the background of SWN was of intermediate difficulty (pairwise Students t-tests: SWN vs BN: t2t = -7.7, df = 60, p < 0.0001; modulated-SWN vs BN: t2t = -6.6, df = 22, p < 0.0001; modulated-SWN vs SWN: t2t = -2.1, df = 28, p = 0.052). Cueing was done with two successive lists and we examined whether there was also any difference between the three noise backgrounds in the change in performance across these two lists. Despite the differences in baseline performance as a function of noise type, the change in performance across the two cueing lists was similar in all three cue-trained groups and was similar to that seen across the first two lists in the Uncued group (in which the two first two lists were the two Learning-task lists) (Fig. 1b; 1-way ANOVA of all four groups: F3, 77 = 1.97, p =0.13; ANOVA of only the three cued groups: F2, 57 = 2.12 p =0.13). Thus, despite context-dependent differences in starting position, the change in SRTs across the first two speech-in-noise lists was the same for all three noise contexts. (Similar effects in the cueing sessions were also seen in groups in which the delay between training and learning tasks was 4 hours or 1 day, discussed below). The consequences of the Cueing were examined on three parameters of the Learning task given 30 mins later: initial (baseline) performance in the 1 learning list, and the early learning (changes in SRTs from list 1 to list 2), and the late learning (changes in SRTs from list 3 to list 6) in the Learning task. Compared to the Uncued group cueing caused a context-dependent immediate improvement in start performance in the Learning task: baseline performance in the Learning task (Fig 1c) was significantly better in the group cued in the same BN (informational masker) context but was not different in either of the other two groups cued in a noise context different from that in the Learning task (1-way ANOVA: F3, 77 = 4.65, p = 0.005; Students t-tests for all pairwise comparisons between BN-cued vs SWN-cued or modulated-SWN-cued groups vs Uncued group: p always < 0.02 and generally p < 0.01; for all pairwise comparisons between SWN-cued or modulated-SWN-cued vs Uncued: p always > 0.15). Thus the BNcued group started out in the Learning task performing slightly better than any of the other three groups. Cueing had no effects on the early phase of learning in the Learning task (Fig. 1d) and the change in SRTs between the first and second Learning lists was the same in all four groups whether or not pretrained in the Cueing task (1-way ANOVA: F3, 77 = 0.33, p = 0.8). However Cueing very significantly reduced the late learning component (Fig. 1e; 1-way ANOVA: F3, 77 = 9.49, p < 0.0001) and this generalized across all three cueing contexts so that equal suppression was found on late learning regardless of the noise masker in the Cueing task (Students t-tests for all pairwise comparisons between Uncued and BN-cued or SWN-cued or modulated-SWN-cued groups: p always < 0.001 and generally p < 0.0001; for all for all pairwise comparisons between the three cued groups: p always > 0.22 and generally p > 0.44). These effects also show that the cueing-evoked interference is not due to methodological factors between Cueing and Learning conditions per se (e.g., sentence differences, voicing differences, or participants’ difficulties in adjusting to noise context differences between Cueing and Learning sessions): as noted in Methods, the sentences used in both tasks were very similar in perceptual characteristics, they were always said in the same female voice in the same tone, and cueing effects were obtained across three different maskers which were identical to the babble noise in the Learning task or only had the same spectral content, or had the same spectral content with the same overall temporal structure. Time course of cueing interference varies with different noise maskers The time course of the cueing effects was examined by increasing the interval between the Cueing and Learning tasks to 4 hours or 1 day, in different groups. Again, for each delay condition, groups could be cued in one of the three different noise types, before testing in the Learning task in BN. The results in the Cueing sessions in these groups were not different from those detailed above for the groups with 30 mins delay between Cueing and Learning tasks, and are not discussed further. The effects of increased delay between Cueing and Learning tasks were assessed on the same three parameters in the Learning task as discussed above, namely, baseline performance, early learning, and late learning, and are compared between the cued groups and the Uncued group (the latter being the same as before). With a 4-h delay after Cueing there was now no longer any effect on baseline performance which was similar in all four groups (Fig. 2a; BN-cued, n = 18; SWN-cued, n = 17; or modulated-SWN-cued groups, n = 18; Uncued, n = 21, same group as before; 1-way ANOVA: F3, 70= 2.34, p = 0.08). There was also no difference in early learning across the four groups (Fig. 2b; 1-way ANOVA: F3, 70= 2.59, p = 0.059). However, as with the 30 -min delay group, there was still a significant difference between groups for the late learning (Fig. 2c; 1-way ANOVA: F3, 70= 10.14, p < 0.0001). In all three cued groups, the amount of learning was significantly less than in the Uncued group (Students t-tests for all pairwise comparisons between BN-cued/SWN-cued/modulated-SWN-cued groups vs Uncued group: p always < 0.0005 and generally p < 0.0001; for all pairwise comparisons between the three cued groups: p always > 0.13). With a 1-d delay after cueing there was also no effect on baseline performance (Fig. 2d; 1-way ANOVA: F3, 62= 0.76, p < 0.52) or in early learning (Fig. 2e; 1-way ANOVA: F3, 62= 0.16, p = 0.92) across the four groups (Uncued; BN-cued, n = 15; SWN-cued, n = 16; or modulated-SWN-cued groups, n = 14). However, there was still a significant difference between groups for the amount of late learning (Fig. 2f; 1-way ANOVA: F3, 62= 7.99, p < 0.0001). In all three Cued groups, the amount was significantly less than in the Uncued group. However there was some recovery in late learning in the group cued in the same context as in the Learning task: the amount of learning in the BN-cued group was much larger than in BNcued groups with the shorter delays or in the other two cued groups with the 1-d delay, and only just significantly different from that in the Uncued group (Students t-tests for all pairwise comparisons between SWN-cued or modulated-SWN-cued groups vs Uncued group: p always < 0.0004, for pairwise comparisons between BN-cued and Uncued group: p = 0.047; for all pairwise comparisons between the three cued groups: p always > 0.055). Thus, again there was a context-independent interference of cueing on late but not early learning, it was weaker in the group cued in the same masker context as in the Learning task, but was as strong in the other two cued groups as with shorter delays. Since all three cued groups here had a night’s sleep interposed between Cueing and Learning, the weakening of effects in the group cued in the same context but not in the two groups cued in different noise contexts does not support a simple model of sleep remediating interference effects of Cueing. Instead they indicate that the three training contexts exert interference effects that decay at different rates and that the effects from the training context most similar to the learning context decay fastest. This effect was partly supported by the time course of recovery of the late learning component (Fig. 2g): recovery patterns were similar for the groups pre-trained in either the SWN masker or the BN masker but very different for the groups pretrained in the modulated-SWN (SWN-BNenv). This hypothesis was directly examined in two groups in which cueing was done with a single list, in either BN or SWN, 1 day before the Learning task. The effect of this single Cueing session on the Learning task are illustrated in Figure 3 where they are compared to the effects seen with training with two lists 1 day before. Cueing with the single list in BN had no significant effects on baseline performance (a1), early learning (b1) or late learning (c1) in the Learning task, compared to the Uncued group or to the group trained with 2 lists in BN the day before the Learning task. Note, as described above, that in the latter Cued group, cueing effects were found only on the late learning component and this was smaller than the effects on this component when cueing was done 1 day previously with 2 lists in SWN or modulated-SWN. Cueing with the single list in SWN 1 day prior to the Learning task resulted in no significant effects on either baseline (Fig. 3a2) or early learning (Fig. 3b2) (effects on these two components of the Learning task were also absent with training with 2 lists in SWN 1 day prior to the Learning task). However, there was still a significant reduction in the late learning component (Fig. 3c2; 1-way ANOVA: F2, 54= 8.58, p = 0.0006); there were no significant differences in effects on this component whether cueing was done with 1 or 2 training lists (Students t-test: t2t = -1.7, df = 34, p = 0.09) and both groups differed from the Uncued group (1-Cueing-list group vs Uncued group: t2t = 2.44, df = 39, p = 0.019; 2-Cueing-list group vs Uncued group: t2t = 4.18, df = 34, p = 0.0002). This effect is consistent with our hypothesis that different contexts in the Cueing condition lead to interference that dissipates with different time courses. DISCUSSION Previous studies of interference with perceptual learning show it acts to prevent learning from occurring at all, when temporal jitter is introduced within a session involving multiple learning stimuli , or blocks learning consolidation when another training session, containing stimuli slightly variant from the originally trained stimuli or contains temporally-jittered stimuli, is applied within 1-4 hours post-initial training 22, . In at least the cases of interference with learning consolidation it can be remediated by an overnight sleep or even a short nap . In both forms of interference, there is specificity to the blocking effect. Interference of learning involving a single learning stimulus is specific to the trained stimulus (e.g., retinal location specificity or task specificity). In the case of the multiple learning stimuli, interference occurs with introduction of temporal jitter between the stimuli to be learnt and can be obtained with temporal jitter in the presentation of two stimuli that are similar but is eliminated by using two stimuli that are less similar, again indicating specificity of action. The present study shows a very different form of interference with perceptual learning: a prospective cued interference that generalizes across context to affect the late learning of a generalized skill. This interference is achieved by a cueing session in the same task (albeit with different stimuli) applied before the Learning training sessions and generalizes so that it can be obtained with cueing contexts different from that used in the subsequent learning. It interferes with the late learning of the general skill of signal extraction from noise (noting that all cueing and learning sessions always involved different sentences), and persists even 1 day after the cueing. The interference could be evoked even by a very transient single cueing exposure 1 day previously. Cueing did possess one stimulus specific element: post-training improvement in the baseline in the Learning task was found only for training in the same noise type. Learning in our speech-in-noise task is achieved very rapidly, both within and across training blocks. Within-block performance asymptotes within the 15 sentences of each list/block 19, , a very much shorter number of trials per block than usually required in many other perceptual learning studies employing a single stimulus in perceptual tasks that are more remote from direct daily experience and therefore less ecologically familiar. A recent study of perceptual learning of multiple visual contrasts in each learning session found that a block of 8-30 trials of each of four interleaved stimulus standards was sufficient to produce maximum perceptual learning . A number of factors work to promote rapid asymptotic within-block performance in the speech-in-noise task: each 4-6 word sentence in our task effectively presents multiple training opportunities in speech-in-noise discrimination and identification, the task is an ecologically-familiar one undertaken everyday by most of us and thus little learning would be needed by the participants of task requirements and criteria, and the specific speech material consists of simple, natural conversational sentences which provide a number of contextual cues and also likely involves top-down influences (such as word and sentence contexts, segmentation cues, lexical knowledge, expectations, higher-level feedback, gaining experience with the talker’s voice, etc; e.g., 42, 43) that all promote perception of connected speech, especially in noise. These factors may also account for the fact that the task also shows rapid across-session learning: across nearly 400 participants 19, , across-session learning consists of an initial rapid phase, between the first and second sessions, and a subsequent slower phase from the second to later sessions and learning appears to asymptote within six sessions 19, 20 (and unpublished data). Wagener and colleagues found an improvement of up to 2 dB in the first two tests of their German speech-in-noise tests, with “no further strong training effect after this initial training” (21, p.146). The very long time course of interference here (interference was present even 1 day after a single cueing exposure) argues against simple adaptation effects as suggested for some forms of perceptual interference (see 41 re 23) and the generalization across noise context to affect the general skill of speech extraction from noise argues against interference at low-level feature representations such as acoustic and phonetic representations. Studies in visual perceptual learning have proposed that perceptual learning, especially of complex but ecologically-commonplace perceptions, is mediated by immediate access to high-level abstract neural representations and only under special circumstances to low-level feature representations (reverse hierarchy theory, RHT ). The perception of words in noise also appears to be very well fitted only by the predictions of the RHT rather than other models , consistent with the strong role of top-down influences on speech-in-noise perception . Low-level processing is gradual and cannot be achieved on a trial-by-trial basis whereas perception and perceptual learning via RHT is very rapid 41, 44, ; this could very well account for the rapid learning in the ecologically-familiar speech-in-noise task especially with the specific type of sentence material used here. Perception through immediate access to high-level representations is argued to be relatively crude 45 and this is inconsistent with the sophisticated processing required to understand speech in noise. However, as the simple, conversational sentences in the speech-in-noise task map readily to natural and familiar situations, perceptual learning in this task may promote access to the high-level representations in a way that allows more sophisticated and generalized learning (since improvements in the task represent improved general ability to extract novel sentences from noise rather than the specific sentences used here). Cued interference was always specific to the late component in this generalized learning across time scales of 30 mins – 1 day, arguing against it being a slowly-developing phenomenon. Further, the fact that cueing caused a context-specific improvement in baseline in the Learning task indicates that it was capable of having – and, in this instance, did have – more immediate effects. Recent studies of auditory perceptual learning indicate that even the early phase of learning consists of substantial amounts of perceptual learning 6, 41 and that early and late perceptual learning differ in the important regard that the former generalizes to untrained stimuli whereas late learning does not . Our data indicates another very important difference between these components with early learning being free from perceptual interference in our task whereas a generalizing, time-decaying forward interference affects late perceptual learning. The fact that this interference was evoked by very different cueing contexts argues for a highlevel locus to the interference, providing another example of top-down influences in the perception of natural speech in noise. In this study the term Cueing has been applied to the initial exposure to the speech-in-noise task to differentiate the effects here from the “Eureka” priming effect 41, . There a single long exposure to an easy instance of a task allows perceptual learning in a later hard instance of the same type of task where learning would not occur otherwise. It is unknown if these opposing end-consequences on perceptual learning are due to the use of simple training stimuli in the Eureka effect and stimuli with significantly more complex relationships here: as noted by Ahissar and Hochstein 41 it is unknown if the same learning rules apply in the two conditions. Speech-in-noise audiometry has a special place in the evaluation of the practical daily consequences of hearing loss and the utility of a specific rehabilitative device in a patient, as it attempts to realistically model the conditions under which we undertake our most important everyday task of communication. Our study carries implications for the use of such tools multiply in the same patients: determining the extent to which prior testing (cueing) in the task is likely to influence the effects seen in a subsequent test session, and the way in which contextual differences between the prior experience and the current test influence the outcomes in the test situation, is important to properly evaluating the effect of