Retained capacity for perceptual learning of degraded speech in primary progressive aphasia and Alzheimer’s disease

Deficits of speech perception have been documented in the three major ‘language-led’ dementias (non-fluent variant primary progressive aphasia [nfvPPA], semantic variant PPA [svPPA] and logopenic variant primary progressive aphasia [lvPPA]) [1‐5] as well as in typical Alzheimer’s disease (tAD) [6‐9]. However, the factors that affect speech perception have been less well studied than language output in these diseases. Normal speech perception entails high-fidelity encoding of incoming acoustic data, parsing of speech from extraneous noises and integration with prior expectations [10]. Healthy listeners rapidly and automatically adapt to speech degradation under challenging listening conditions, based on prior auditory experience [11]. This reflects perceptual learning: improved accuracy of perceptual processing following sustained exposure to the auditory stimulus, modulated by ‘top-down’ predictive mechanisms [12, 13].

Perception of degraded speech is impaired in patients with nfvPPA, svPPA and tAD [14, 15], whereas implicit auditory sequence learning and disambiguation of degraded speech are retained in nfvPPA [10, 16], albeit with reduced flexibility in using contextual information. Processing degraded sensory stimuli taxes the functional integrity of cortical circuits [17, 18], and disorders such as PPA and tAD constitute test cases for exposing such effects because they strike auditory processing networks early and relatively selectively. In contrast, perceptual learning reflects functional brain reorganisation or plasticity; it might therefore help compensate for effects of neurodegeneration [19]. Adaptation to degraded speech engages areas (such as sensorimotor cortex) beyond classical language and auditory networks [20, 21] and may be enhanced by cholinergic stimulation [15]. Taken together, these considerations suggest that understanding and perceptual learning of degraded speech might constitute a powerful and sensitive probe of neural network integrity and residual plasticity in neurodegenerative syndromes.

In the present study, we assessed identification of degraded speech and associated auditory perceptual learning in patients with PPA and tAD relative to healthy older individuals. We used sine-wave speech as a model paradigm. Sine-wave speech is a radical perceptual alteration that reduces speech signals to a series of ‘whistles’ that correspond to formant contours, retaining the long-range temporal scaffold of speech but stripped of all spectral detail (examples are available in the additional files). Sine-wave transformation renders speech initially unintelligible, yet induces spontaneous perceptual learning in healthy listeners primed to its linguistic origin [11, 22]. This effect relies on ‘top-down’ perceptual integration of apparently dissimilar acoustic events into a coherent speech-like signal. It is therefore likely a priori to depend on cognitive operations that are instantiated across the language network. To explore the effect of semantic predictability on the ‘top-down’ disambiguation of degraded speech [10], we applied sine-wave manipulations to a verbal category with uniformly high predictability (numbers) and a verbal category in which predictability could be varied on the basis of familiarity (geographical place names). These semantic categories are relatively preserved in svPPA [2, 23], allowing predictive processing mechanisms to be distinguished from semantic disintegration. Neuroanatomical associations of sine-wave speech processing in the patient cohort were assessed using voxel-based morphometry (VBM).

On the basis of available evidence [3‐5, 10, 15, 24‐27], we hypothesised that PPA syndromes and tAD would be associated with differential impairment of sine-wave speech identification and perceptual learning relative to controls. We predicted that nfvPPA and lvPPA would be associated with more severe perceptual decoding deficits than other groups, but that all groups would show retained adaptation to degraded speech and increased reliance on prior predictability, particularly in svPPA. Drawing on neuroanatomical evidence in the healthy brain and PPA [10, 13, 20, 28‐30], we further hypothesised that sine-wave speech perception deficits would correlate with grey matter loss in posterior superior temporal and inferior parietal cortices, whereas modulation by prior predictability and perceptual learning effects would correlate with anterior sensorimotor, prefrontal and anterior temporal grey matter.

Background demographic, neuropsychological and clinical data for all participant groups are presented in Table 1; group performance profiles on the experimental tasks are presented in Table 2.

In the present study, we have demonstrated deficits of degraded speech perception in major syndromes of PPA and tAD relative to healthy control participants. Syndromic groups were stratified by their overall accuracy in decoding sine-wave speech, with patients with nfvPPA and lvPPA exhibiting the most severe and consistent impairments. The findings in tAD extend previous evidence in this disease [15]. Syndromic profiles were modulated by the prior predictability of verbal content: The svPPA and tAD groups showed a significant advantage for perceiving sine-wave speech with highly predictable content (numbers) compared with less predictable content (place names), whereas within the geographical condition, all groups showed a performance advantage for more familiar (‘near’) over less familiar (‘far’) place names, and this advantage was exaggerated in patient groups compared with the healthy control participants. All syndromic groups exhibited some capacity for auditory perceptual learning following sustained exposure to sine-wave speech: this occurred spontaneously, and there was evidence that the effect generalised to ‘new’ as well as ‘trained’ speech tokens. However, patients with svPPA and tAD showed a perceptual learning effect only for relatively predictable verbal content. These effects were evident after adjusting for performance on natural speech perception and geographical knowledge tasks and therefore not attributable to more generic deficits of phonological or semantic processing in the patient groups. Taken together, the present results imply a degree of residual cerebral plasticity in these syndromes, manifesting as resilient auditory perceptual learning of degraded speech.

Our finding that decoding of degraded speech is impaired in the nfvPPA, lvPPA and tAD groups corroborates recent evidence for core auditory processing deficits in these syndromes [5‐8, 10, 14, 27, 30]. The less uniform decoding deficit identified in the svPPA group in the present study was also anticipated on the basis of previous work; in the healthy brain, decoding of degraded speech engages ‘top-down’ (including semantic) mechanisms that disambiguate the speech stream based on prior predictability [10, 13, 37]. Efficient access to stored semantic ‘priors’ (including ready access to lower-frequency priors) when interpreting degraded speech is likely to become increasingly limiting as verbal content becomes less predictable [26]; less predictable verbal content would place increased demands on semantic processing resources. This would account both for the striking performance advantage for (highly predictable) sine-wave numbers over (less predictable) place names shown in this study by patients with svPPA and for the ‘echo’ of this condition discrepancy effect in the performance advantage for more familiar over less familiar place names shown by all participant groups. This was not simply the consequence of a dwindling semantic lexicon; it was observed after taking geographical semantic competence into account, in keeping with a more specific limitation on the recruitment of semantic mechanisms during predictive processing of speech signals. Furthermore, the tAD group, but not the lvPPA group, showed a significant condition discrepancy effect, suggesting that these Alzheimer variant syndromes may be characterised by separable pathophysiological mechanisms [2].

Participant groups did not differ in perceptual learning of sine-wave numbers. Indeed, the magnitude of the perceptual learning effect across patient groups (including those with marked overall deficits of degraded speech perception) was comparable to that of healthy older control participants (Table 2, Figs. 2 and 3). Moreover, most individual patients in each syndromic group showed perceptual learning effects within the healthy control range. There was no association between general cognitive performance (indexed by MMSE and WASI matrices scores) and perceptual learning index for numbers or place names. Together, these findings argue for dissociable physiological mechanisms mediating the accuracy of degraded speech decoding and adaptation based on sustained, unsupervised exposure to degraded speech, and they suggest that perceptual learning of degraded speech may be relatively resilient to the effects of background cognitive decline. The syndromic profiles in the present study further suggest that perceptual learning is modulated by prior verbal predictability (the svPPA and tAD groups showed the effect for strongly predictable but not less predictable verbal stimuli), in line with current models of degraded speech learning based on minimising prediction errors [13]. Although data in PPA are limited, our findings are consistent with previous work in nfvPPA, indicating that separable mechanisms underpin decoding of sensory detail in degraded speech stimuli, ‘top-down’ predictions about such stimuli and implicit learning based on auditory experience [10, 16]. There are precedents for a dissociation of sensory accuracy and sensory learning or plasticity in other disorders (e.g., developmental dyslexia and amblyopia [38]), with potential substrates at cognitive, neurophysiological and neuroanatomical levels [39]. Speech may be a particularly potent stimulus to expose the component mechanisms of a processing hierarchy. Whereas perceptual decision making on speech and other complex auditory stimuli typically rests on integration of multiple spectrotemporal features, perceptual learning may depend on the extraction of more specific lower-level properties from ‘bottom-up’ sensory data, honed by ‘top-down’ predictions based on prior auditory experience and used in turn to update those predictions [13, 39]. This reciprocal interaction between sensory traffic and predictions could be instantiated on different neuroanatomical scales, ranging from local cortical circuits to large-scale distributed brain networks that could be differentially disrupted by neurodegenerative proteinopathies [10, 17, 40].

The neuroanatomical correlates of degraded speech decoding accuracy and perceptual learning identified in our patient cohort support the behavioural evidence that these processes are at least partly dissociable. Identification of both sine-wave numbers and place names was associated with grey matter volume in left angular gyrus, a region affected by the neurodegenerative pathologies studied here [41, 42] and previously implicated in processing speech under challenging listening conditions in functional neuroimaging and virtual lesion studies in the healthy brain [17, 43‐46]. The present evidence in a patient cohort with variably impaired perception of degraded speech corroborates this previous work and further suggests that integrity of angular gyrus plays a critical role in determining whether degraded speech is disambiguated successfully. However, this region acts as the hub of a distributed processing network. Its functional connectivity and interactions with other modes of the network may be modulated by a number of factors, including output task, semantic context and perceived intelligibility [47]. In line with this, we identified additional neuroanatomical correlates that may mediate the effect of altered verbal predictability of degraded speech (summarised in Table 3). Grey matter in the left planum temporale was correlated with identification of sine-wave place names but not numbers at the prescribed threshold. This region is engaged in parsing the auditory scene under conditions of high computational load [13, 28, 36].

Candidate loci for ‘top-down’ control of perceptual analysis under different conditions of verbal predictability were identified in the left inferior frontal gyrus (for more highly predictable, sine-wave numbers) and left temporal pole (for less predictable sine-wave place names). Our findings do not resolve the mechanisms by which these regions communicate during sine-wave speech decoding, which are likely to differ between dementia syndromes. Disambiguation of degraded speech based on relatively constrained predictive algorithms (such as word verification or number identification) is likely to engage top-down control mechanisms in the inferior frontal cortex, a region that is heavily involved in nfvPPA [10], whereas decoding of ‘novel’ linguistic environments with less predictable verbal content (such as degraded place names) may demand active computation of the ‘best fit’ between incoming speech signal statistics and stored verbal concepts, accessed via the anterior temporal lobe semantic network that is blighted in svPPA [14, 26, 30, 48]. Considered together, our findings suggest that the overall accuracy of degraded speech decoding in these syndromes is likely to depend on a distributed peri-Sylvian network closely overlapping classical language cortices. Both nfvPPA and lvPPA (and less consistently svPPA) have been shown to be associated with atrophy or dysfunction of the dorsal language network, with involvement of anterior and posterior regions extending beyond the zone of maximal atrophy in particular syndromes [2, 5]. Impaired accuracy of sine-wave speech identification might plausibly result from a ‘double-hit’ to inferior frontal and temporoparietal regions previously implicated in decoding speech and other complex auditory signals [4, 5, 30].

A neuroanatomical substrate for perceptual learning of degraded speech (sine-wave numbers) was identified in the inferolateral post-central gyrus. This sector of sensorimotor cortex does not form part of the canonical language network and was not associated with sine-wave speech identification accuracy in this study, consistent with dissociable neural mechanisms for perceptual decision making and perceptual learning of degraded speech. However, this sensorimotor region hosts cortical representations of lips, mouth and tongue that are engaged during speech perception, particularly under difficult listening conditions in which subvocal rehearsal may help to resolve ambiguous speech sounds [13, 20] or in the context of disease processes primarily affecting the auditory cortex [21]. Sensorimotor cortices are relatively spared in PPA syndromes and tAD [1, 49]; we propose that a critical determinant of perceptual learning capacity (cerebral plasticity) is the degree of atrophy (or relative preservation) of these areas. Our findings suggest that some degree of residual neural plasticity is maintained beyond vulnerable language and auditory networks across canonical dementia syndromes. A stronger claim would be that some form of compensatory functional enhancement drives perceptual learning in the face of neurodegenerative pathology. We cannot evaluate this claim on the basis of the present evidence, though we note that no patient group in our study showed increased perceptual learning capacity relative to healthy older control participants. To understand the nature of the observed effects fully will require functional neuroimaging approaches that can address network connectivity and activity changes directly. Functional neuroimaging paradigms based on degraded speech stimuli have been developed in the healthy brain [13, 50], but they have yet to be applied to patients with dementia. The present findings corroborate our previous psychoacoustic work in tAD [15] and the recent demonstration that patients with nfvPPA benefit from retraining strategies for speech production and fluency, with lasting and generalisable improvement of communication function [51]. Our findings raise the further intriguing possibility that the efficacy of such communication interventions may be enhanced by engaging perceptual learning capacity (i.e., residual cerebral plasticity). If we are to exploit this potential, quantitative behavioural and neuroanatomical markers of perceptual learning will be required. Sine-wave speech (and related degraded speech manipulations) may offer a convenient and well-established route to development of relevant plasticity biomarkers.

This study suggests a number of directions for future work. Sine-wave speech served in the present study as a model paradigm for understanding speech under challenging listening conditions. We found that capacity for perceptual learning of this radically degraded speech-like signal is retained across diverse dementia syndromes, despite variably impaired understanding of the signal. Measures of individual responses (Fig. 3) suggest that these stimuli may represent novel markers for assessing and tracking communication function in particular patients and could have therapeutic potential. Dynamic markers of this kind might stratify dementia syndromes but also transcend conventional syndromic boundaries, constituting ‘stress tests’ of speech processing and auditory scene analysis in earlier-stage PPA and tAD and also presenting a target for intervention. This need not await the advent of disease-modifying therapies; combining currently available symptomatic pharmacotherapies (such as cholinesterase inhibitors) with speech retraining might be one rational approach [15, 16]. However, more information is required about the diagnostic sensitivity, specificity and relevance of perceptual measures on degraded speech over the course of disease, based on replication of these findings in larger patient cohorts, correlation with indices of daily life communication functions, and extension to other speech manipulations. In addition, detailed understanding of the pathophysiological mechanisms of degraded speech processing and perceptual learning in neurodegenerative syndromes will require functional neuroimaging and connectivity-based techniques that can capture activity profiles and time-varying interactions between brain regions.

This work has broad neurobiological and clinical implications. Neurobiologically, the findings suggest that neurodegenerative proteinopathies dissect dissociable mechanisms for auditory pattern decoding and adaptation and expose the critical brain substrates for these processes. Clinically, this work forecasts a fresh emphasis on dynamic physiological capacity and functional plasticity in dementia that should motivate novel biomarker development and neurorehabilitation strategies.