Short-term memory capacity and sensitivity to language statistics in dyslexia and among musicians.

Poor short-term memory (STM) capacity in individuals with dyslexia (IDDs) and enhanced STM capacity in musicians are well documented, yet their causes are disputed. Previous studies also found poor use of stimuli statistics by IDDs and enhanced use by musicians. We hypothesized that these observations are functionally related, as follows: Enhanced sensitivity to statistics facilitates musicians' benefit from each exposure, and reduced sensitivity to statistics hinders IDDs' benefit. Thus, larger group differences are expected for larger exposure: STM capacity, which is sensitive to item familiarity, will thus be larger among musicians, and smaller among IDDS, particularly for high-frequency items. Testing this hypothesis using syllable span, we found that musicians' advantage and IDDs' difficulty were indeed larger for high-frequency syllables than for low-frequency ones. By contrast, benefits from sequence repetition did not differ between musicians, controls and IDDs, suggesting that online sequence learning is based on a different mechanism. To test this dissociation we recruited, in addition to native Hebrew speakers, native English speakers, matched to the Hebrew-speaking controls. Their spans for high-frequency syllables in Hebrew, which do not have high frequency in English, were small - as expected from reduced exposure to these syllables. Yet, their benefit from sequence repetition was similar to that of the three Hebrew-speaking groups. Taken together, these experiments suggest that different sensitivities to item frequency explain some of the population-related variability in STM tasks. By contrast, benefits from sequence repetition do not depend on item familiarity, and do not differ between groups.

IDDs' STM capacity and utilization of implicit prior knowledge are both reduced

Developmental dyslexia is a specific difficulty in accurate or fluent word recognition, poor decoding, and poor spelling abilities (American Psychiatric Association, 2013) in spite of adequate hearing levels, normal intelligence and adequate educational opportunities. Beyond reading difficulties, individuals with dyslexia (IDDs) suffer from substantial difficulties in phonological awareness and in span tasks, in which a participant is asked to reproduce a sequence of items in the order of their presentation (Jeffries and Everatt, 2004; Roodenrys and Stokes, 2001; Snowling et al., 1986; Snowling, 1981).

Traditional theories of dyslexia attribute these difficulties to a core deficit in phonological representations (e.g., Snowling, 2000) or to reduced efficiency in accessing these representations (Ramus, 2014; Ramus and Szenkovits, 2008). However, more recent theories, such as the “anchoring deficit hypothesis” (Ahissar, 2007; Ahissar et al., 2006; Oganian and Ahissar, 2012), have proposed a learning-related core deficit in dyslexia: reduced sensitivity to stimulus statistics in spite of adequate sensory processes (Jaffe-Dax et al., 2015). Specifically, IDDs' neural traces decay faster (Jaffe-Dax et al., 2017; Jaffe-Dax et al., 2018; Lieder et al., 2019), yielding a functional “leakage” of memory traces, which reduces the effective rate of acquiring familiarity and hence results in poorer long-term representations (Banai and Ahissar, 2017).

The implications of the anchoring-deficit hypothesis partially overlap with those of the phonological account: phonological representations are also expected to be somewhat impoverished as part of a slower rate of acquisition of stimuli statistics. However, the predictions of the anchoring-deficit hypothesis are not specific to phonology, and in some cases they even oppose the predictions of the phonological account. Specifically, based on the phonological-deficit theory, difficulties of IDDs are expected to increase with phonological complexity, and perhaps to decrease with familiarity, due to a compensation for the phonological deficit with multiple presentations and unimpaired benefits of regularity. If benefits from regularity are not assumed to be part of the core deficit in dyslexia, but rather IDDs are assumed to suffer from a persistent perceptual noise (Sperling et al., 2005) or a fixed impairment in processing of sounds (Pasquini et al., 2007), then multiple exposures to the same stimuli (e.g., a word or a syllable) are expected to increase the signal-to-noise ratio. Hence, the quality of representations of frequent items for IDDs will increase, and that will result in a relatively smaller difficulties for frequent, compared with infrequent, items.

Learning-related theories, as opposed to those for perception-based deficits, predict that relative difficulties should increase with stimulus-specific exposure. In line with this prediction, several studies have reported deficits associated with linguistic regularities, such as morphological deficits (Rispens et al., 2004; Schiff and Ravid, 2007). For example, adult IDDs have a reduced morphological benefit in acquiring new words (Kimel and Ahissar, 2019).

Thus, the hypothesis of reduced accumulation of language statistics yields the following prediction: relative difficulties of individuals with dyslexia will increase with exposure due to the slower learning rate of regularities and of repetitions in the input, as illustrated in Fig. 1 (a shallower slope of the red vs. blue lines). Given evidence that item frequency bears a substantial influence on performance in span tasks (e.g., Hulme et al., 1997; Nimmo and Roodenrys, 2002; see details below), we now ask whether the hypothesis of IDDs' reduced sensitivity to language statistics provides a valid alternative account to the common observation of IDDs' reduced STM.

Fig. 1

A schematic illustration ofthe expected performance in span tasks as a function of exposure to the tested stimuli in 3 populations with different sensitivities to stimuli statistics. All participants start from scratch (zero performance) and improve with no saturation (e.g., log function). Group difference is expected to increase with exposure and to be larger for high-frequency items (vertical dashed line on the right) as compared with low-frequency items (vertical dashed line on the left). To test the impact of item frequency, we measured spans of two syllable types: frequent (large amount of exposure) and infrequent, respectively.

Overall, the observation of poor scores in span tasks among IDDs is very robust (though see Wimmer, 1993). STM in IDDs is typically assessed with the standard Digit Span test (e.g., Oganian and Ahissar, 2012; Spring 1976). Performance in Digit Span is correlated with reading speed within and across languages (Naveh-Benjamin and Ayres, 1986), suggesting functional relations between poor span scores and poor reading skills. Studies in which single non-words or lists of non-words were used, also yielded poor STM scores (e.g., Jeffries and Everatt, 2004; Roodenrys and Stokes, 2001; Snowling et al., 1986; Snowling, 1981).

Poor span scores among IDDs have traditionally been attributed to these individuals' poor phonological representations (e.g., Snowling, 2000) or to their reduced efficiency in accessing these representations (Ramus, 2014; Ramus and Szenkovits, 2008). However, we propose a different source for this amply reported deficit among IDDs, which is grounded in a well-known dependency between STM and long-term memory (LTM).

The effective capacity of STM is sensitive to item frequency

The capacity of our STM is limited, ranging from the classic estimation of 7 items (“magical number 7 ± 2”, Miller, 1956) to the more recent suggestion of 4 items (“magical 4 ± 1”, Cowan, 2010). The most common procedure for measuring STM is the Span Task, and many studies have found that item characteristics greatly affect performance in these tasks. For example, spans of frequent words are larger than those of infrequent ones (Hulme et al., 1997; Roodenrys et al., 2002), and span for words is greater than that for non-words (Hulme et al., 1991). The benefit of familiarity also applies at the syllabic level - typically assessed with a slightly different task: non-word repetition (e.g., Gathercole and Adams, 1993 in two- and three-year-old children; Gathercole and Baddeley, 1989; Gathercole et al., 1999). For example, Nimmo and Roodenrys (Nimmo and Roodenrys, 2002) showed that non-word repetition accuracy and recall of syllable sequences are both higher for syllables that occur more frequently in polysyllabic English words. Similarly, Tremblay and colleagues found greater accuracy in non-word-repetition with higher frequency of their first syllable (Tremblay et al., 2016). Importantly, these studies corroborate recent imaging results suggesting that STM does not encompass representations separate from long-term memory and that, instead, it reflects access to long-term memory (Sreenivasan, Curtis, & D'Esposito, 2014). Sreenivasan and colleagues claim that sensory cortex keeps high-fidelity representations of Working Memory content. For example, they review a number of studies that showed that actively maintained visual objects are encoded by patterns in the occipital and temporal cortex regions, which are specialized for object representations. Moreover, in a recent review Cowan (2019) analyses both behavioral and imaging studies and summarizes that there is no need to assume a separate “STM copy”, but rather STM is the portion of LTM that is currently activated.

An alternative explanation to IDDs' reduced STM

Given that high-frequency items yield larger spans in the general population (e.g., Nimmo and Roodenrys, 2002), we hypothesized that a smaller benefit from exposure would hinder the STM spans of IDDs, and particularly so for spans of frequent items, as plotted in Fig. 1. The items used in span tasks are usually frequent language items, and although it is well known that long-term item frequency has a significant impact on performance in span tasks, a systematic study of this factor's effect on IDDs had never previously been undertaken. We therefore manipulated item frequency and tested the prediction that IDDs' spans would be particularly poor (compared with controls matched for age and reasoning abilities) when high-frequency items are used.

Musicians have an elevated STM capacity

There is ample evidence that musical proficiency is correlated with performance in STM tasks (Chen, Penhune and Zatorre, 2008b, 2008a; Janata and Grafton, 2003; Janata et al., 2002). For example, children and adults who have received musical training outperform non-musicians on digit, word and non-word memory tasks (Franklin et al., 2008; Fujioka et al., 2006; Lee et al., 2007; Parbery-Clark et al., 2009, Chandrasekaran and Kraus, 2010). The majority of these tasks are conducted with stimuli such as letters, digits and other highly familiar items. However, items' familiarity was not taken into account as an independent factor that might have special effect on musicians, although enhanced sensitivity of musicians to stimuli statistics was previously suggested (Shook et al., 2013). Based on this, we hypothesized that musicians' enhanced STM relies to some extent on these individuals' enhanced learning rate for stimulus statistics, which in turn facilitates their performance in a related task, as shown in Fig. 1 (purple line). Namely, we hypothesized that musicians' performance in STM tasks mirrors that of IDDs due to the same underlying mechanism. Indeed, musicians' reading skills are enhanced as compared with controls matched for age and reasoning abilities (Weiss et al., 2014).

Sensitivity to long-term stimulus statistics is expected to affect performance in span tasks

Fig. 1 summarizes our hypothesis, which attributes different respective learning rates to controls, IDDs, and musicians, and hence predicts that performance following long-term acquisition will particularly differ between these groups for high-frequency items. The main underlying principle is that cumulative benefits from encounters with stimuli are affected by the benefit from each exposure. In the case of IDDs, this benefit is reduced due to faster memory decay, leading to a decreased learning rate, whereas for musicians the benefit is higher and their learning rate is accordingly increased.

Crucially, though learning rate decreases with exposure, it does not saturate, and improvement continues for years. This continuous learning has been shown across domains (the exponential/power law of practice – Crossman, 1959; Heathcote et al., 2000), and it applies among other things to reading. For example, even after years of reading in a second language, individuals' reading rate in that language does not reach their native-language reading rate. Similarly, although the reading rate of IDDs continuously improves with practice, they typically do not catch up with their peers. This also applies to spans: spans for words and syllables in one's second language are not expected to be as large as those of words in one's native language. We therefore predicted that group differences would increase with familiarity (right-hand side of Fig. 1, compared with the left-hand side). Comparing with controls with similar exposure, then, IDDs' STM will be particularly poor for frequent items whereas musicians' STM will be particularly high.

We chose to focus on syllables for a few reasons: their familiarity is not strongly dependent on reading experience, they usually do not carry semantic content in Hebrew, and there is evidence of their presence as separate mental entities quite early in development (e.g., Perfetti et al., 1987). Thus, illiterate adults and nursery-school children both show similar success rates in manipulating syllables, but only literate individuals successfully operate with phonemes (Liberman et al., 1974; José Morais, Cary, Alegria and Bertelson, 1979). For frequent syllables we chose Consonant-Vowel (CV) syllables, and for infrequent ones we chose Vowel-Consonant (VC) syllables. CV syllables are part of the syllabic vocabulary in almost all known languages (Sommer, 1970). They are acquired early and predominate both in babbling and in early meaningful speech (Stoel-Gammon, 1989). In Hebrew they constitute the majority of syllables (Ben-Dror et al., 1995). Contrary to CV syllables, VC syllables are rare in languages generally (Clements and Keyser, 1983), and in Hebrew they are an order of magnitude less frequent than CV syllables (Ben-Dror et al., 1995; discussed in Share and Blum, 2005).

We measured spans of frequent (CV) and infrequent (VC) syllables in three populations: IDDs, controls and musicians. Based on previous studies, mean spans were expected to be higher for frequent syllables than for infrequent ones. Overall performance of IDDs was expected to be lower than that of the control group, and musicians' performance was expected to be higher than that of the controls. Crucially, we predicted a previously unattested interaction: that syllabic frequency would benefit musicians more than it would controls, and that IDDs would benefit less than controls.

In addition to the benefit from item frequency, span scores also benefit from sequence repetition. Some studies (Bogaerts et al., 2015; Szmalec et al., 2011; though see Henderson and Warmington, 2017; Staels & Van den Broeck, 2014, 2015; Wang et al., 2016) have found that IDDs benefit less from sequence repetition, and this might be related to IDDs' reduced spans, and perhaps also to musicians' increased spans. We therefore also assessed benefits from sequence repetition in the same task.

Method

General cognitive assessments

Non-verbal intelligence was assessed using the Block Design task (a subtest from the Hebrew version of the Wechsler Adult Intelligence Scale, WAIS-III; Wechsler, 1997). The Block Design task measures spatial reasoning abilities, and is often used to match groups on non-verbal reasoning.

Standard short-term memory skills were assessed with the sub-tests of Digit Span (Wechsler, 1997): Digit Forward and Digit Backward. Digit Forward requires immediate oral repetition of orally presented sequences of digits, and Digit Backward requires immediate oral repetition of orally presented sequences of digits in reversed order.

Reading measures

We used three measures of reading proficiency: single-word reading, pseudo-word reading, and paragraph reading. The lists of real words and pseudo-words were standard lists (Deutsch and Bentin, 1996) and were presented with diacritics, which make Hebrew orthography transparent. Assessment of reading in context was performed through the reading of a four-paragraph academic-level text in Hebrew (standardized for students by our lab; Ben-Yehudah et al., 2001). Participants were instructed to read the text aloud as quickly and accurately as possible, yet slowly enough for them to be able to subsequently answer a simple content question, so as to encourage text comprehension. Both accuracy and rate (i.e., the total number of words read in 1 min) were scored for all reading tests.

Stimuli – estimating syllabic frequency

The exact frequency of syllables in Hebrew has not been calculated, probably because syllables cannot be automatically parsed from written Hebrew without vowel diacritics. We therefore assessed syllable frequency by calculating the distribution of syllables in the largest publicly available corpus of spoken Hebrew that has a phonological transcript (~6,500 word tokens; The Corpus of Spoken Israeli Hebrew – CoSIH). In this corpus about 54% of the syllables are CV, 25% are CVC, and only 5.5% are VC. The mean (SD) % of occurrence in the corpus for the specific syllables that were used in the current study is 0.58 (0.69) for the CV syllables and 0.06 (0.11) for the VC syllables, t = 6.8, p < 10−8. Table S1 in the supplementary material shows the count of each syllable.

Procedure

We used a Syllable Span protocol (Oganian and Ahissar, 2012; Weiss et al., 2014), which is based on the Digit Span protocol (WAIS-III; Wechsler, 1997), and administered it four times: with frequent (CV) and infrequent (VC) syllables, and with and without series repetition. VC syllables were constructed by switching the vowel and the consonant phoneme in each CV syllable, e.g., ko and ok. In the no-repetition condition, no syllable was repeated throughout the experiment. In the repeated condition we used two sequences that were gradually elongated and were presented in an interleaved manner. For example, the four initial sequences for the CV repeated condition were: 1. /ve//tsu/2. /ko//ʃa/3. /ve//tsu//pi/4. /ko//ʃa//ħu/; and for the non-repeated condition: 1. /tsu//na/2. /ʃe//ko/3. /li//ʃu//ga/4. /di//bu//mo/.

The VC spans were always administered before the CV spans in order to avoid a possible benefit for the VC syllables due to training. The order of repetition vs. no-repetition was counterbalanced across participants. A recording of a native female Hebrew speaker was used, and syllables were played at a fixed interval of 1 s (SOA). The score for each condition is the number of sequences that the participant reproduced correctly.

Participants

Participants were recruited through advertisements posted at the Hebrew University campuses, at two other colleges in Jerusalem, and at the Jerusalem Academy of Music and Dance. Participants were paid for their participation. Candidate participants filled in a detailed questionnaire about their formal academic education, their history of reading difficulties (including any previous diagnoses), their musical background and their medical condition. Exclusion criteria were: hearing problems, psychiatric medications other than attention deficit medication, and a formal musical background in the case of control-group candidates and IDDs (that is, more than two years of either playing a tonal instrument or of formal singing education). All participants received all of their schooling in Israel. Individuals who passed this preliminary screening were invited to a basic assessment session of cognitive and reading skills. On the basis of their performance in this session, one participant with dyslexia and one control participant were excluded due to below-average cognitive scores (i.e., Block Design score < 7; Wechsler, 1997). One IDD had accurate (100% correct) non-word reading, and was therefore excluded from the study. The dyslexic group included eight participants who typically take medication for ADHD (e.g., Concerta). Based on our previous studies (Ben-Yehudah et al., 2001) they were not excluded from participation, but they did not take this medication on test days. The majority of the musician group overlaps with the group in the study of Weiss and colleagues (Weiss et al., 2014), but the data reported in the current study had not been previously reported elsewhere.

Thirty-four control individuals, 35 IDDs, and 37 musicians took part in Experiment 1. All members of the musician cohort were either students at the Jerusalem Academy of Music and Dance or professional musicians with an academic degree in music. Data of participants whose score was less than 3 in any of the span conditions were removed from the results, as we have attributed their poor performance to lack of understanding of the task (n = 1); the 3 oldest control participants and the 7 youngest musicians were excluded in order to match age with the other groups. Thus, we report results of 31 controls, 34 IDDs and 30 musicians matched for age and general reasoning abilities, as shown in Table 1. As expected, there were significant differences among the groups in reading and reading-related measures. In addition, consistent with previous findings (e.g., Ackerman et al., 1990; Torgesen et al., 1990), IDDs' scores were significantly lower than those of controls and musicians in the Digit Span task.

Table 1

Participant Characteristics: Hebrew-Speaking Controls, IDDs, and Musicians.

Measure	Controls	IDDs	Musicians	Significant Group difference
	N = 31 (13 M)	N = 34 (11 M)	N = 30 (9 M)
Years of musical practice	Up to 2	Up to 2	12.7 (3.7)
Age (years)	25.16 (2.2)	24.32 (3.1)	24.27 (2.2)	n.s.
Block design (scaled)	12.71 (3.0)	12.91 (2.7)	13.73 (2.6)	n.s.
Digit Span (scaled)	11.32 (2.8)	7.88 (1.9)	12.13 (3.0)	Controls-IDDsMusicians-IDDs
Single-words reading accuracy (% correct)	96.91 (4.2)	87.25 (8.0)	99.17 (1.7)	Controls-IDDsMusicians-IDDs
Single-words reading rate (words/minute)	96.6 (29.5)	69.27 (25.0)	119.71 (22.5)	All pairwise comparisons
Pseudo-words reading accuracy (% correct)	90.46 (11.2)	60.78 (18.5)	96.53 (3.6)	Controls-IDDsMusicians-IDDs
Pseudo-words reading rate (words/minute)	57.82 (24.2)	33.67 (11.9)	76.26 (13.9)	All pairwise comparisons
Paragraph reading accuracy (% correct)	98.65 (1.4)	95.02 (4.3)	98.85 (1.1)	Controls-IDDsMusicians-IDDs
Paragraph reading rate (words/minute)	138.83 (21.0)	98.69 (22.6)	134.09 (15.9)	Controls-IDDsMusicians-IDDs

Means (SD) are presented. M – Male participants.

Means for groups denoted in the right column significantly differ at the p < .05 level after Bonferroni correction.

Results

The number of correctly reproduced sequences in each condition (score; Table 2) was analyzed using a mixed-design analysis of variance (ANOVA) with Syllable Type (frequent vs. infrequent) and Repetition (repeated vs. non-repeated) as within-subject factors, and Group (controls, IDDs, and musicians) as a between-subject factor. Effect sizes for group comparisons are calculated using Cohen's d (Cohen, 2013).

Table 2

Syllable Span Scores.

Type of syllable	Type of sequence	Measure	Controls	IDDs	Musicians
Frequent	Non-repeated	Score	8.4 (1.5)	7.1 (1.5)	10.6 (2.3)
		Span	5.6 (1.0)	5.0 (0.9)	7.2 (1.4)
	Repeated	Score	10.6 (2.2)	9.1 (2.2)	13.2 (2.1)
		Span	7.1 (1.2)	6.3 (1.2)	8.3 (1.0)
Infrequent	Non-repeated	Score	5.9 (1.1)	5.0 (1.2)	6.5 (1.2)
		Span	4.2 (0.6)	3.9 (0.7)	4.8 (0.8)
	Repeated	Score	5.5 (1.2)	5.3 (1.3)	6.6 (2.3)
		Span	4.3 (0.7)	4.2 (0.8)	4.9 (1.3)

Scores and spans (mean (SD)) of the Syllable Span task with frequent and rare syllables, with and without sequence repetition. The score for each sequence was either 0 or 1: 0 if there was at least one mistake/missing syllable, and 1 if the recall was perfect. The total score for each subject for each condition is the sum of the sequences' scores. The span is the length of the longest sequence that the participant was able to recall correctly.

The main effects were as expected: musicians achieved higher scores than controls did, and controls achieved higher scores than did IDDs (main effect of Group, F(2,92) = 33.18, p < 10−10, η2 = 0.419; a post hoc analysis with the Bonferroni correction: controls vs. IDDs: p < .013, Cohen's d = 0.43; controls vs. musicians: p < 10, Cohen's d = 0.60; Fig. 2; Table 2). Scores for frequent syllables were higher than for infrequent ones (main effect of Syllable Type, F(1,92) = 762.3, p < 10−45, η2 = 0.892; Fig. 2) in each of the groups (controls: mean-difference = 3.82, SE = 0.26, p < 10−25, η2 = 0.706; IDDs: mean-difference = 3.0, SE = 0.25, p < 10−20, η2 = 0.619; musicians: mean-difference = 5.37, SE = 0.26, p < 10−35, η2 = 0.821).

Fig. 2

Scores of infrequent (VC) vs. frequent (CV) syllable spans in the three test groups: IDDs (red), controls (blue), and musicians (violet). Groups are plotted (left to right) according to their increasing benefits from syllabic frequency. Spans are larger for frequent syllables than for infrequent ones (long-term frequency effect). Circles denote individual scores. Star symbols denote means; error bars denote 1 SEM; horizontal bars denote medians. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

In line with our hypothesis, musicians benefited from syllable frequency more than controls did, and controls benefited more than IDDs (Syllable Type × Group interaction, F = 22.2, p < 10, η2 = 0.325; a post hoc analysis on the difference between scores for frequent and for infrequent syllables with Bonferroni correction: controls vs. IDDs: p < .068, Cohen's d = 0.62, controls vs. musicians: p < 10, Cohen's d = 1.00; Fig. 2). Additionally, group difference was larger for frequent (CV) than for infrequent (VC) syllables (CV syllables – controls vs. IDDs: mean-difference = 1.35, SE = 0.43, p < .007, Cohen's d = 0.84, musicians vs. controls: mean-difference = 2.45, SE = 0.45, p < 10−6, Cohen's d = 1.36, musicians vs. IDDs: mean-difference = 3.80, SE = 0.44, p < 10−12, Cohen's d = 2.15; VC syllables – controls vs. IDDs: mean-difference = 0.53, SE = 0.29, p = .203, Cohen's d = 0.51, musicians vs. controls: mean-difference = 0.91, SE = 0.29, p < .009, Cohen's d = 0.75, musicians vs. IDDs: mean-difference = 1.43, SE = 0.29, p < 10−5, Cohen's d = 1.20; Bonferroni corrected for multiple comparisons).

Repeated sequences yielded higher scores than non-repeated sequences (main effect of Repetition, F(1,92) = 67.6, p < 10−11, η2 = 0.424; Fig. 3; Table 2) in each of the groups (controls: mean-difference = 0.89, SE = 0.24, p < 10−3, η2 = 0.424; IDDs: mean-difference = 1.15, SE = 0.23, p < 10−5, η2 = 0.213; musicians: mean-difference = 1.37, SE = 0.25, p < 10−6, η2 = 0.253). But in contrast to the group difference in benefits from syllable frequency, the three groups did not significantly differ from one another in their benefits from sequence repetition (Repetition × Group interaction: F(2,92) = 0.98, p = .380, η2 = 0.021; Fig. 3).

Fig. 3

Difference between scores for spans with repeated series(Rep.)and spans without repeated series(No rep.)in sequences of infrequent (VC, left) and frequent (CV, right) syllables, in the three groups: IDDs (red), controls (blue), and musicians (violet). Benefit from series repetition does not significantly differ among the three populations, and is larger for CV syllables (right) than for VC syllables (left). Circles denote individual scores. Star symbols denote means; error bars denote 1 SEM; horizontal bars denote medians. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

The benefit from repetition was larger for the frequent CV than it was for infrequent VC syllables (Syllable Type × Repetition interaction: F(1,92) = 67.1, p < 10−11, η2 = 0.422). In fact, repetition effect was significant for the frequent syllables (mean-difference = 2.26, SE = 0.20, p < 10−18, η2 = 0.581), but not for the infrequent syllables (mean-difference = 0.01, SE = 0.19, p = .947, η2 = 0). Overall, the pattern of the benefit from repetition as a function of syllable frequency did not significantly differ among the groups (Syllable Type X Repetition × Group interaction: F (2,92) = 1.25, p < .291, η2 = 0.027; controls: benefit from repetition for frequent syllables mean-difference = 2.19, SE = 0.35, p < 10−7, η2 = 0.301, and for infrequent mean-difference = −0.42, SE = 0.33, p = .207, η2 = 0.017; IDDs: benefit from repetition for frequent syllables mean-difference = 1.97, SE = 0.33, p < 10−7, η2 = 0.276, and for infrequent mean-difference = 0.32, SE = 0.32, p = .307, η2 = 0.011; musicians: benefit from repetition for frequent syllables mean-difference = 2.60, SE = 0.36, p < 10−10, η2 = 0.369, and for infrequent mean-difference = 0.13, SE = 0.34, p = .692, η2 = 0.002).

Discussion

Overall the three groups showed the expected larger spans for lists composed of frequent syllables (CV) vs. those composed of infrequent (VC) syllables (we could not assess the correlation between the individual syllable frequency and performance beyond the level of syllable type due to the limitations imposed by a span paradigm). Musicians' overall syllable span scores were higher than controls', whose scores were higher than those of the IDDs. In line with our hypothesis, IDDs' gain from syllable frequency was smaller than that of the controls, and the musicians' gain was greater than that of the controls. These results – taken together with the typical use of frequent stimuli in standard STM assessments – suggest that the reports on musicians' enhanced STM, and IDDs' reduced STM, stem to an extent from musicians' enhanced use of long-term statistics and IDDs' reduced use thereof. By contrast, the benefits of IDDs and musicians from sequence repetition did not significantly differ from those of controls, suggesting that the sensitivity to series repetition is similar among the groups and hence cannot explain the difference in spans. For dyslexia, these results are in line with previous findings about adequate benefit from series repetition (Staels & Van den Broeck, 2014, 2015; Wang et al., 2016; though see Bogaerts et al., 2015).

Musicians' performance was also significantly better than controls' for the infrequent syllables (and not just for the frequent syllables). There are two potential explanations to this difference. The first is that STM of musicians is generally superior to that of controls, and thus their performance is better in both conditions. The second is that the amount of exposure to the relatively infrequent syllables was sufficient for significantly improving musicians' performance. Fig. 1 illustrates this point: group difference is smaller on the left side than on the right side, but we do not know a-priori where VC syllables are on this graph. Importantly, benefits from series repetition do not stem from the same mechanism as benefits from long-term frequency of single syllables (Kalm et al., 2013; Schwarzlose et al., 2008). Thus, although repetition of VC series was not found to be beneficial in this paradigm, we still suggest that musicians' enhanced sensitivity to item frequency enhances their VC spans compared to controls (and increases musicians' CV spans even more). Also, repetition of VC series was found to be beneficial when more repetitions were given (Kimel et al., 2020).

The differing benefits from item frequency seen among these groups, and the similarity in benefit that they all derive from sequence repetition, suggest a dissociation between benefit from long-term item frequency and the learning of a repeated sequence. However, in all three groups the benefit from sequence repetition was larger for CV sequences than it was for VC sequences. This difference seems puzzling: if benefit from series repetition does not depend on item familiarity, as suggested from the lack of differences among groups in serial-order repetition, what other difference underlies the larger advantage of CVs compared with VCs in sequence repetition? A likely candidate is the easiness of sequencing. CVs are more fluently co-articulated. It is easier to join CV syllables (compared to VC syllables) into sequences during production, thus increasing the effective STM (Miller, 1956). More formally, CVs are categorized as “light syllables”, whereas VCs are categorized as “heavy syllables” (Duanmu, 2010; Lunden, 2011) based on their different ending type (rime). Heavy syllables attract stress (the Weight-Stress Principle, Duanmu, 2010; Velupillai, 2012), and a word (or a sequence of syllables) can only have one primary stress. Therefore, a sequence of VCs tends to be perceived and articulated with a stress on each syllable, supporting segmentation rather than chunking, since stressed syllables are treated as word onsets (Cutler and Norris, 1988). By contrast, a sequence of CVs can be perceived and articulated as a long multi-syllabic word. Thus, the easiness of sequencing varies for different syllable structures due to their different manner of articulation, rather than their different frequency. Given these inherent differences between CV and VC syllables, we suggest that although a specific set of syllables was used in the current study, these results apply to CV vs. VC syllables in general, since this characteristic is general.

According to our account, the relative easiness of sequencing CV syllables is especially beneficial for learning repeated series of CV syllables, and thus facilitates the effect of series repetition. But this facilitation in repeated series learning is not due to CV syllables being frequent. Learning of repeated sequences of VC syllables is reduced since VC syllables are not easy to sequence, and thus to chunk. Syllable frequency has an effect on the span in general, and this effect is separate from that of syllable easiness of sequencing.

We propose that there is a group difference in sensitivity to the frequency of syllables, but that there is no significant group difference in sensitivity to series repetition. Effective learning of a repeated series, we propose, relies mainly on the easiness of sequencing syllables, and this does not differ between the groups. To test the hypothesis that the benefits from syllable frequency and series repetition are dissociated, we conducted Experiment 2.

For the three groups that participated in Experiment 1, CV syllables are both frequent and relatively easy to sequence whereas VC syllables are both infrequent and difficult to sequence. To disentangle these two factors, we recruited a group with reduced familiarity with CV syllables. Thus, for this group, CV syllables are less frequent but should still be easy to sequence if fluency of sequencing is (largely) an exposure-independent property of CV syllables. For this group we therefore predicted reduced CV spans, but similar benefits from series repetition.

Experiment 2

As previously mentioned, CV syllables constitute the majority of syllables in Hebrew (Ben-Dror et al., 1995). But their prevalence in English is substantially lower than in Hebrew (28%–38%, Dauer, 1983). Based on this substantial difference, we hypothesized that native English speakers with only basic familiarity with Hebrew would show reduced CV spans as compared with native Hebrew speakers (as illustrated in Fig. 1 – left side vs. right side). We further hypothesized that, in spite of these individuals' reduced spans for CV syllables, their benefits from series repetition of these CV syllables would be similar to those of native Hebrew speakers. We hypothesized this since the easiness of sequencing, which underlies the main benefits from series repetition, is based not on syllable familiarity but on their manner of articulation. We reasoned that such a dissociation would directly indicate that for series repetition, the benefits thereof are independent of the familiarity with the composing items, thus contrasting it with STM's dependence on familiarity with the test items.

VC syllables are more rare than CV syllables in all languages and specifically in English (Clements and Keyser, 1983; Dauer, 1983). Thus, their reduced frequency and reduced easiness of sequencing overlap also for English speakers, similarly to Hebrew speakers (the three groups that took part in Experiment 1). Thus, only CV syllables are appropriate for assessing the dissociation between frequency and easiness of sequencing in English speakers, and consequently only CV syllables were used in Experiment 2.

Native English speakers without reading difficulties were recruited through advertisements posted at the school of international students in Hebrew University. All recruited participants had taken Hebrew as a second-language class, and their familiarity with Hebrew was at a basic level. We applied similar inclusion criteria to those used for the Hebrew-speaker controls (e.g., no learning difficulties and musical education of only up to two years). The data of the youngest English-speaking participant were removed in order to match age across the groups. Thus, the control group of Experiment 1 and the group of native English speakers were matched for age, general reasoning abilities, and scaled Digit Span scores in their native language (Table 3).

Table 3

Participant Characteristics: English-Speaking Controls.

Measure	Hebrew-speaking Controls	English-speakingControls	t-value
	N = 31 (13 M)	N = 29 (15 M)
Age (years)	25.2 (2.2)	23.6 (2.2)	t = 1.55 ˄
Block design (scaled)	12.7 (3.0)	11.6 (2.6)	t = 1.64 ˄
Native language Digit Span (scaled)	11.3 (2.8)	11.6 (3.0)	t = −0.43 ˄

Means (SD) are presented. M – Male participants. ˄ – n.s.

Data of participants whose score was <3 in one of the span conditions were removed from the analyses due to our attribution of their poor performance to misunderstanding of the task; data of one English-speaking participant were consequently excluded. Participants for whom the difference between the conditions with and without repetition was smaller than their group mean by more than 2.5 standard deviations were defined as outliers, and were removed from the analysis; one additional English-speaking participant was consequently excluded (for this participant the difference was −4). Thus, we report the results of 29 English-speaking controls.

We used the same procedure as we did in Experiment 1, but only with CV syllables: with and without sequence repetition (Oganian and Ahissar, 2012; Weiss et al., 2014). The order of repetition vs. no-repetition was counterbalanced across participants. In order to employ English-like syllables, a recording of a native female English speaker with native-English articulation was used.

The number of correctly reproduced sequences in each condition (score) was analyzed using a mixed-design analysis of variance (ANOVA) with Repetition (repeated vs. non-repeated) as a within-subject factors, and Group (controls, IDDs, musicians, and English-speaking controls) as a between-subject factor. For Hebrew-speaking controls, IDDs and musicians, a subset of the data collected in Experiment 1 (CV syllables only) is reanalyzed in Experiment 2.

As expected, scores of English-speaking controls for the CV syllables were lower than those of Hebrew-speaking controls despite their matched Digit Span scores in their native language (Table 3). Their CV spans were even lower than those of Hebrew-speaking IDDs (main effect of Group: F(3,120) = 55.73, p < 10−21, η2 = 0.582; Bonferroni Post Hoc tests: IDDs vs. English-speaking controls: mean-difference = 1.51, SE = 0.42, p < .003, Cohen's d = 1.02; Fig. 4). As expected, scores for repeated sequences were higher than for non-repeated ones (main effect of Repetition: F(1,120) = 154.17, p < 10−22, η2 = 0.562; Fig. 4). However, there was no difference in the effect of repetition between the groups (Repetition X Group: F(3,120) = 1.76, p = .159, η2 = 0.042; Bonferroni Post Hoc tests were all non-significant, the comparison with the biggest mean difference was musicians vs. English-speaking controls: mean-difference = 1.08, SE = 0.48, p = .160, Cohen's d = 0.60; Fig. 4).

Fig. 4

Spans of CV syllables, with and without series repetition: 4 populations with different degrees of exposure to CV syllables (Hebrew vs. English speakers) and different sensitivities to item frequency (IDDs, controls, and musicians) differ in absolute scores but not in the benefits from series repetitions.

As predicted, native English speakers, who are less familiar with CVs, had smaller CV spans than native Hebrew speakers did - long-term experience with single items influences span. Yet English speakers did not significantly differ from Hebrew speakers in their benefit from series repetition of these syllables. These results support the hypothesis that the benefits of series repetition do not substantially rely on item frequency (beyond basic familiarity): the frequency of CV syllables is lower for native English speakers, and still their benefit from sequence repetition is comparable with that seen in native Hebrew speakers.

General discussion

In Experiment 1, we measured spans for infrequent (VC) syllables and compared them with spans for frequent (CV) syllables, with and without sequence repetition in three populations: Hebrew-speaking IDDs, controls, and musicians. IDDs benefited less from syllable frequency than controls did, and musicians benefited more than did the controls. By contrast, the benefit in span for repeated series did not significantly differ among the groups. To further separate the factors underlying the benefit of long-term familiarity from the factors underlying the benefit of sequence repetition, we conducted Experiment 2. Native English speakers, who are less familiar with CV syllables, were administered the CV span task with and without sequence repetition. Their scores were very poor, as expected. Yet, their benefit from sequence repetition was similar to that seen in Hebrew speakers.

Group difference in the benefit derived from syllable frequency sheds light on the mechanisms underlying the reduced spans of IDDs on the one hand, and the enhanced spans of musicians on the other. Span tasks are usually performed using frequent items, so it is likely that reports on the reduced performance of IDDs, and the enhanced performance of musicians, in fact rely to a large extent on reduced/enhanced benefit from item frequency and not on reduced/enhanced STM in general.

By contrast, we found no significant difference among the groups in the benefit derived from sequence repetition. In the literature, the impact of sequence repetition on IDDs' performance is disputed (see Majerus and Cowan, 2016 for a review). For example, a Hebb repetition learning paradigm (Hebb, 1961), in which a participant is exposed to a repeated sequence of syllables interleaved with non-repeated sequences, has been administered to adult IDDs in various studies. Some studies have reported that sequence learning poses a unique difficulty in dyslexia (Bogaerts et al., 2015; Szmalec et al., 2011), but others failed to replicate this observation despite using the same stimuli and analyses (Staels & Van den Broeck, 2014, 2015).

Evidence from brain studies is consistent with our finding as it suggests a dissociation between the mechanisms that underlie learning of distributional statistics and those that underlie learning of repeated sequences, and hence a reduced benefit from frequency is not necessarily coupled with a reduced benefit from sequence repetition. Learning of a repeated sequence relies on the hippocampus (Kalm et al., 2013; Kok and Turk-Browne, 2018; Krishnan et al., 2016), and this learning is impaired when hippocampal structures are impaired (Schapiro et al., 2014; though see Gagnon et al., 2004). But categories of single items (distributional statistics) are represented semantically (Schwarzlose et al., 2008) and perceptually (Alain et al., 2007; Garrido et al., 2009) in specific cortical regions.

It has been previously suggested that musicians' enhanced STM might rely on enhanced general chunking skills that are developed as part of the musical training (Talamini et al., 2017). However, a recent meta-analysis showed that musical training does not transfer and does not predict other cognitive abilities (Sala and Gobet, 2017), in line with studies that showed limited transfer between training in the general population (Banai and Ahissar, 2013; Jakoby et al., 2019). Enhanced STM might be associated with innate musical competence or other characteristics of individuals who become musicians (Corrigall et al., 2013). Regardless of the source of the superior performance of musicians in STM tasks (e.g., Franklin et al., 2008; Parbery-Clark et al., 2009), given the lack of group difference in benefits from series repetition, our results do not support a generally enhanced chunking mechanism utilized by musicians in span tasks. Rather, we propose that musicians' enhanced sensitivity to long-term item frequency underlies their increased spans.

Experiment 2 dissociates the effect of long-term item frequency and learning of repeated sequences for controls (i.e., individuals without dyslexia that do not have extensive musical training). However, it does not necessarily entail that sensitivity to item frequency explains the enhanced performance of musicians and the reduced performance of IDDs, as other factors than what we found for controls might be influential for musicians and IDDs.

An alternative explanation for musicians' pattern of results could be that musicians' STM is larger in general, regardless of stimuli type. This is consistent with musicians' better performance for the infrequent (VC) syllables in Experiment 1. While group difference is smaller in the infrequent condition, musicians' performance is still generally superior compared to the other groups, for both syllable types.

While the enhanced-sensitivity account does not directly predict a significant group difference in the infrequent condition, all Hebrew-speaking participant have had some experience with VC syllables, which are all part of the words in their native language (though rare). Thus, if musicians' benefit from exposure is enhanced, they might have reached a significant advantage compared to controls for the infrequent condition too. Importantly, group difference in the frequent condition is higher, as predicted and illustrated in Fig. 1. By contrast, the generally-elevated-STM account needs an additional assumption to explain why musicians' CV spans are especially large (Group X Syllable type interaction). Perhaps CV syllables are more readily represented in STM among musicians (and less readily represented for IDDs). If we do not assume enhanced sensitivity to exposure among musicians, we need to assume a different cause for this effect, whose nature should be studied.

Impaired reading is part of the definition of dyslexia, and musicians' enhanced reading skills have been characterized before (e.g., early reading in children Anvari et al., 2002; Weiss et al., 2014). Higher reading rate of single words and pseudo-words among musicians was also found in this study (Table 1). Similarly to the benefit from long-term knowledge in STM tasks, reading benefits from long-term language regularities and repetitions (Carlisle, 2000; Mahony et al., 2000), and it has been shown that IDDs benefit to a lesser extent from such regularities in reading (Kimel and Ahissar, 2019). We showed an enhanced benefit of musicians from long-term syllable frequency. This elevated utilization of accumulated statistics might also underlie musicians' enhanced reading skills. Importantly, the ability to segment and analyze phonetic units improves with reading, but for CV syllables it is relatively high even with lower reading proficiency, unlike for other syllable types (Morais, Alegria and Content, 1987). This is an advantage of our design, since differences in reading proficiency cannot explain the bigger group difference for CV syllables. Rather, spans are a tool that assesses more basic mechanisms, which are shared with reading proficiency.

The reduced benefit of IDDs from syllable frequency, manifested as a bigger group difference in the frequent condition, supports learning-based accounts of dyslexia, and specifically the anchoring-deficit hypothesis (Ahissar, 2007; Ahissar et al., 2006). Namely, IDDs' relative difficulties increase with exposure (i.e., CV syllables) rather than with phonological complexity (i.e., VC syllables), consistent with our prediction, as illustrated in Fig. 1. This is further supported by numerous studies that find an STM deficit in dyslexia when frequent items are used (e.g., Jeffries and Everatt, 2004; Roodenrys and Stokes, 2001) and reduced usage of other language statistics, such as morphological structure, in dyslexia (Kimel and Ahissar, 2019; Schiff and Ravid, 2007).

Taken together, our results propose that musicians possess skills mirroring those of IDDs: enhanced vs. reduced accumulation of long-term stimuli statistics, respectively. This explains their different STM characteristics, and perhaps also differences in reading skills. Based on data from 4 populations: Hebrew-speaking IDDs, controls, and musicians, and English-speaking controls, the current study is the first to suggest that unique STM characteristics of musicians and IDDs can be explained by a single underlying mechanism: benefit from long-term item frequency.

Credits

Eva Kimel: Conceptualization, Formal analysis, Investigation, Writing - Original Draft, Writing - Review & Editing, Visualization. Atalia Hai Weis: Conceptualization, Methodology, Investigation, Writing - Original Draft. Hilla Jakoby: Methodology, Investigation. Luba Daikhin: Investigation, Writing - Original Draft, Writing - Review & Editing. Merav Ahissar: Conceptualization, Writing - Original Draft, Writing - Review & Editing, Supervision, Funding acquisition.