Decoding of Speech Information using EEG in Children with Dyslexia: Less Accurate Low-Frequency Representations of Speech, Not “Noisy” Representations

The amplitude envelope of speech carries crucial low-frequency acoustic information that assists linguistic decoding. The sensory-neural Temporal Sampling (TS) theory of developmental dyslexia proposes atypical encoding of speech envelope information <10 Hz, leading to atypical phonological representations. Here a backward linear TRF model and story listening were employed to estimate the speech information encoded in the electroencephalogram in the canonical delta, theta and alpha bands by 9-year-old children with and without dyslexia. TRF decoding accuracy provided an estimate of how faithfully the children’s brains encoded low-frequency envelope information. Between-group analyses showed that the children with dyslexia exhibited impaired reconstruction of speech information in the delta band. However, when the quality of speech encoding for each child was estimated using child-by-child decoding models, then the dyslexic children did not differ from controls. This suggests that children with dyslexia encode neither “noisy” nor “normal” representations of the speech signal, but different representations.


Introduction 62
Dyslexia is a disorder of development. Classically, a child has shown apparently typical 63 language acquisition and cognitive development until faced with the task of learning to read. 64 Suddenly, the affected child shows a specific problem with reading and spelling that cannot be 65 accounted for by low intelligence, poor educational opportunities, or obvious sensory or 66 neurological damage. Reading is the process of understanding speech when it is written down. 67 To acquire early reading skills, the child must learn the visual code used by their culture for 68 representing speech as a series of symbols. Logically, individual differences in acquiring 69 reading could be related developmentally to either spoken language processing or visual code 70 processing or both (Snowling, 2000;Stein and Walsh, 1997). However, it is important to 71 recognise that the visual code that must be learned is not a neutral visual stimulus. It is a 72 culturally-specific code that is taught directly using symbol-sound correspondences, such as 73 the alphabetic correspondences used to represent English phonology (the sound system of the 74 language). Learning this visual code typically begins a few years into the development of a 75 spoken language system, and it is now widely recognised that individual differences in 76 children's pre-reading phonological awareness (their ability orally to recognise or manipulate 77 phonological units in words such as syllables, rhymes or phonemes) is a causal determinant of 78 how readily their visual code learning progresses (Ziegler and Goswami, 2005, for cross-79 language review). Indeed classically, 'deficits' in the representation and use of phonological 80 information have been seen as critical in the etiology of developmental dyslexia (Catts, 1993;81 Stackhouse and Wells, 1997; Stanovich and Siegel, 1994;Swan and Goswami, 1997; 82 Snowling, 2000). 83 To understand the nature of this 'phonological deficit' in dyslexia, longitudinal studies 84 that begin with infants are required. Evidence from both electrophysiological and behavioural 85 measures suggests that infants at family risk (genetic risk) for dyslexia already exhibit deficits 86 when a task requires lots of phonological information to be stored in short-term memory, or a 112 task requires speeded or repeated retrievals of phonological information, or it requires the 113 extraction of speech information from noise (Ramus and Svenkovits, 2008). To our knowledge, 114 however, neural studies have not yet been able to contrast the "noisy representations" view of 115 the phonological 'deficit' in dyslexia with the opposing view of Ramus and Svenkovits that 116 the phonological representations developed by the dyslexic brain are "normal". We describe 117 an initial neural approach to exploring this contrast here. 118 In the last 5 years, neural studies with children with dyslexia have been applying novel 119 computational methods to EEG and MEG data to try to measure the quality of children's 120 phonological representations directly. These methods either use original speech features such 121 as the envelope to predict the brain's neuroelectric responses (forward models, from speech 122 envelope to brain response, Di Liberto et al., 2018), or reconstruct speech-based representations 123 of envelope information from the brain's neuroelectric responses to speech input (backward or 124 speech reconstruction models, from brain response to speech signal, Power  These prior TRF studies also included younger children matched for reading level to 136 the children with dyslexia. This is an important experimental control for the effects of learning 137 on the developing brain. It is known that learning to read in itself changes phonological 138 representations, as orthographic knowledge gained through reading affects performance in oral 139 listening tasks for both children and adults (Ehri and Wilce, 1980;Ziegler and Ferrand, 1998). 140 In principle, younger reading-level-matched (RL match) children provide a control for the 141 effects of reading experience on the developing language system. Both of the prior EEG studies 142 using both backward and forward TRF models reported significantly greater stimulus 143 reconstruction accuracy (Power et al., 2016) or neural response accuracy (Di Liberto et al., 144 2018) regarding low-frequency acoustic speech information for their RL controls compared to 145 their dyslexic participants. Accordingly, in EEG studies the dyslexic brain shows less accurate 146 speech encoding of low-frequency envelope information than the brains of younger children 147 matched for reading experience, suggesting a fundamental difference in encoding certain 148 aspects of linguistic information. The exception was the MEG study using speech-in-noise 149 tasks, and the authors explained their null result by suggesting that discriminating speech in 150 noise may be facilitated by learning to read (Destoky et al., 2020). If speech-in-noise 151 performance depends on the level of reading attained, then children with dyslexia should 152 perform at the same level as younger children matched for reading, as found by Destoky et al. 153 (2020). 154 To date, therefore, data from neural TRF studies do not suggest that the phonological 155 representations of words developed by children with dyslexia are "normal" regarding the 156 encoding of low-frequency acoustic information. At the same time, however, it is not clear that 157 these representations are "noisy" or "fuzzy" regarding this low-frequency information. Rather, 158 low-frequency acoustic information in the speech signal appears to be represented less 159 accurately in the neural mental lexicons of children with dyslexia. This suggests that speech 160 itself may be perceived differently by children with dyslexia. Rather than being a noisy signal 161 for the dyslexic brain, the signal itself may be represented in an atypical manner, with less 162 accurate representation of low-frequency information ). This reduced accuracy 163 for envelope information may possibly be compensated by an over-weighting of allophonic 164 and other information (Bogliotti et al., 2008), which may be represented with greater specificity 165 by children with dyslexia. It is also possible that both low-frequency envelope information in contrast (ba/wa) to children with dyslexia and age-matched and RL controls, changing "ba" to 172 "wa" either by varying amplitude rise time in synthetic syllables or by varying frequency rise 173 time. The children with dyslexia were significantly poorer at discriminating "ba" from "wa" 174 when the phonetic change depended on amplitude rise time, but were significantly better than 175 both age-matched and RL controls when the phonetic change depended on frequency rise time. would also complicate the process of learning to read, since any visual symbol system that is 181 being learned will have been designed for learners who hear speech differently from children 182 with dyslexia. This would apply whether the visual code is an alphabetic system or any other 183 orthographic system (see . 184 In the current study, we set out to use a stimulus envelope reconstruction approach to 185 begin to tackle these theoretical issues ("noisy" representations versus "normal" 186 representations) with the same data. Our reasoning was as follows. Using a backward TRF 187 modelling approach, thereby using neurophysiological responses to reconstruct speech 188 envelopes < 10Hz, natural speech listening data can be used either to compare the decoding 189 accuracy of the models that are reconstructed between groups of children, or within individual information is encoded less accurately by children with dyslexia in the delta and theta bands. 217 Regarding the "noisy" representations question, given the frequent observation that children 218 with dyslexia appear to show no difficulties in speaking and listening tasks that do not tax 219 phonological knowledge, it may be that the children are operating with perceptually stable 220 phonological representations at the speech envelope level that are not "noisy" for their users. 221 222

Participants 224
Fifty-one children were participated in this study. Twenty-one participants were 225 typically developing children (mean age of 109.3 ± 5.4 months) and thirty participants had 226 developmental dyslexia (mean age of 110.7 ± 5.6 months). The unequal group sizes arose due 227 to Covid-19, which necessitated the cessation of testing part-way through the study, thereby 228 also preventing testing of the recruited RL control group. One of the dyslexic children only 229 completed 5 minutes of the EEG session, due to lots of head movement, not listening to the 230 story, and taking out the earphone phone during data collection. We therefore excluded this 231 child from further analysis. The children with dyslexia did not have any additional learning 232 difficulties (e.g., ADHD, dyspraxia, autistic spectrum disorder, developmental language 233 disorder), and were recruited through learning support teachers. The absence of the additional 234 learning difficulties was confirmed based on school and parental reports and our own 235 behavioural testing. All participants had a nonverbal IQ above 84, and their first language 236 spoken at home was English. All children received a short hearing screen across frequency 237 range of 0.25 -8 kHz (0.25, 0.5, 1, 2, 4, 8 kHz) using an audiometer, and they all were found 238 to be sensitive to sounds within the 20 dB HL range. SES data were not formally collected, but 239 children were attending state schools (equivalent to US public schools) situated in a range of

Behavioural tests 246
A series of standardised tests (see Table 1) of written and spoken language were 247 administered in schools prior to the EEG session to assess cognitive development (please see was administered to assess phonological awareness at the rhyme and phoneme levels, along 258 with rapid naming of objects and digits (RAN). The children's amplitude rise time thresholds 259 were estimated using 3 psychoacoustic threshold tasks based on (1) sine tones; (2) tones made 260 from speech-shaped noise; (3) a synthetic syllable "ba". As shown in Table 1, the children with 261 dyslexia had significantly poorer reading and spelling skills than the control children, 262 significantly poorer phonological skills, and significantly poorer amplitude rise time 263 discrimination in a synthetic syllable task (discriminating differences in rise time of the syllable 264 "ba").

Experimental set-up and stimuli 278
The children were seated in an electrically shielded soundproof room. The auditory 279 stimuli (through earphones) were presented at a sampling rate of 44.1 kHz to the participant 280 while EEG data were collected at a sampling rate of 1 kHz using a 128-channel EEG system 281 (HydroCel Geodesic Sensor Net). The stimuli were natural speech presented as a 10-minute 282 long story for children, The Iron Man: A Children's Story in Five Nights by Ted Hughes. The 283 story was read in child-directed speech and was presented in ten sections, each of which lasted 284 about one minute followed by a 2 second gap. During the experiment, participants were 285 instructed to listen to the speech carefully and to look at a red cross (+) shown on the screen 286 that was in front of them. 287 288

Auditory Stimuli and EEG Data Pre-processing 289
In this study, we first obtained the broad-band envelopes by calculating the absolute 290 value of the analytical signal of the speech stimuli. The broad-band envelopes were then filtered 291 at frequency range of 0.5 -8 Hz to extract the low frequency envelopes. The low frequency 292 envelopes were used for the analyses used here. 293 The collected EEG data were referenced to Cz channel and then band-passed filtered 294 into frequency range of 0.5 -48 Hz using a zero phase FIR filter with low cutoff (−6 dB) of 295 0.25 Hz and high cutoff (−6 dB) of 48.25 Hz (EEGLab Toolbox; Delorme and Makeig, 2004). 296 Bad channels were detected and interpolated through spherical interpolation (EEGLab 297 Toolbox). The EEG data were downsampled to 100 Hz and filtered to extract delta (0.5 -4 298 Hz), theta (4 -8 Hz) and alpha (8 -12 Hz) frequency bands. The data were further 299 downsampled to 50 Hz to reduce the computational costs, and then epoched into the 1-minute 300 story parts. 301 302

Stimulus Reconstruction Using Backward TRF Model 303
To investigate how accurately the children were encoding the low-frequency 304 information in the speech signal, we used a decoding accuracy method based on the backward 305 TRF model (Crosse et al., 2016). Given the focus of TS theory on the envelope, and given that 306 the speech envelope carries the most critical low-frequency acoustic information aiding 307 linguistic decoding, the speech envelope was chosen as the decoding target in our study. The 308 TRF model was used to reconstruct the stimuli envelopes from the EEG signals at each 309 frequency band of interest (delta, theta, alpha). The backward TRF model estimates the 310 coupled with a cross-validation procedure. The validation procedure was the "leave-one-out" 321 cross-validation (using mTRFcrossval function from the mTRF Toobox) in which each trial is 322 "left out" or used for testing and the remainder are used to train the model and this procedure 323 is repeated across all trials. Eight values were used for the ridge parameter (λ = 0.1, 1, …, 10 6 ) 324 in this procedure and the value that gave the highest average correlation score was chosen as 325 the optimal value for the ridge parameter. This optimal value was then used to train the model. 326 The Pearson correlation between the estimated envelope and the actual envelope was 327 used to estimate the performance of the model, with a higher correlation value indicating more 328 accurate decoding by the model. value for each group. This then enabled us to assess whether there were differences in speech 345 decoding accuracy regarding low-frequency envelope information between children with 346 dyslexia and age-matched controls. In essence, the reconstruction accuracy for the typically-347 developing children was utilised as a benchmark to compare with the reconstruction accuracy 348 obtained for the dyslexic children. Figure 1 provides a

Within-Child Backward TRF Analysis 359
To discover whether the perceptual experience of low-frequency envelope information 360 in the speech signal is consistent (or stable) for any individual child, we also built a single 361 backward TRF model for each child in the study based on the data from that child. This within-362 child approach enabled us to assess whether the speech-based representations developed by 363 children with dyslexia are less consistent, "fuzzy" or "noisy" for each frequency band of 364 interest. This was achieved by using 80% of the data (eight story sections) for each child to 365 train the TRF model for that child, and then using the remaining data from that child (two story 366 sections) to test the model. We calculated the Pearson correlation between the estimated speech 367 envelope and the actual speech envelope for each child separately for each frequency band. 368 This resulted in a single correlation score for each band for each child. These scores were then 369 summed by group for each band and compared between groups. Figure 2 provides a schematic 370 diagram of the analytic approach used for the within-child backward TRF analysis. 371 We also applied the Wilcoxon rank sum test to compare the ridge parameter values for 372 the within-child model between the control and dyslexic groups separately for the three 373

Computation of the Chance Level for Backward TRF 385
To check the statistical significance of stimulus reconstruction accuracy as estimated 386 by the backward TRF models (both between-group and within-child), we computed null 387 models for each frequency band of interest. To obtain the null models, we randomly selected 388 ten control children and EEG data were permuted across different story sections for each child. 389 We then built a single model for each child. To achieve this, 80% of the data for each child 390 were used to train the null model for that child, and the remaining data were used to test the 391 null model. The Pearson correlation between the estimated speech envelope and the actual 392 speech envelope for each child separately for each frequency band was then calculated for these 393 null models, resulting in a single correlation value for each band and for each child. We finally 394 computed the mean reconstruction accuracy by averaging across the accuracy scores obtained 395 for the selected children. This was done 100 times for each band, in order to calculate the 396 probability density functions (PDF, a statistical measure that determines the probability 397 distribution of a variable) for the null models and thereby establish chance levels. Note that the 398 area under the PDF curve should be 1, hence for our data the smaller the range of Pearson 399 correlations on the x axis, the higher the value required for the PDF on the y axis. 400

Does the Accuracy of Speech Envelope Decoding Vary Between Control and Dyslexic 403
Children? 404 To investigate whether the accuracy of speech envelope decoding estimated from EEG 405 recorded from the dyslexic brain is different from that recorded from the typically-developing 406 brain, the Between-Group Backward TRF analysis (see Section 2.5) was applied separately for 407 each frequency band. To check the statistical significance of the stimulus reconstruction 408 accuracy obtained for each frequency band by group, we first compared decoding accuracy in 409 each band to decoding accuracy for the null models (the estimate of chance level for each band, 410 see Section 2.7). Statistical significance by band and group is shown in Figure 3. The 411 modelling showed that stimulus-reconstruction accuracy was significantly greater than chance 412 in the delta band only ( Figure 3A). Reconstruction accuracy in the theta and alpha bands was 413 not statistically different from the noise values in these bands for either group (Figure 3B,C).

424
To compare the accuracy of decoding between the control and dyslexic groups in the 425 delta band, we applied the Wilcoxon rank sum test. This showed that reconstruction accuracy 426 for the control group in the delta band as derived from 100 TRF models was significantly 427 greater than reconstruction accuracy for the dyslexic group (Wilcoxon rank sum test, z = 3.72, 428 p = 0.0002). This difference in speech envelope decoding suggests that control brains encode 429 a more accurate neural representation of the speech envelope in the delta band compared to 430 dyslexic brains, as shown in Figure 4.

Information in Speech? 441
To explore whether children with dyslexia have "noisy" or "fuzzy" representations of 442 the low-frequency envelope information in the speech signal, the Within-Child Backward TRF 443 analysis method (see Section 2.6) was employed separately for each frequency band. The 444 statistical significance of the stimulus reconstruction accuracy obtained for each frequency 445 band for the within-child analyses was computed by considering chance level for that band (the 446 null models, see Section 2.7). Comparisons with the null models showed that stimulus-447 reconstruction accuracy was above chance-levels (alpha = 0.05) for both groups in all three 448 frequency bands ( Figure 5).    Accordingly, as shown in Figure 6, the backward TRF model can reconstruct the stimulus 477 envelopes for each individual child in each band consistently from that child's EEG data. 478 Furthermore, this envelope information is reconstructed above chance levels for each 479 individual child in all three bands, in contrast to the reconstruction consistency achieved for 480 the between-group analysis, in which decoding accuracy was only above chance for both 481 groups of children in the delta band condition (Figure 3). The within-child group comparisons 482 ( Figure 6) show that the consistency of decoding accuracy is always above chance, and is not 483 statistically different irrespective of whether the child is a child with dyslexia or a control child. 484 The modelling indicates that the neural representation of the low-frequency envelope 485 information in the speech signal for each child is consistent for that child, providing them with 486 a stable percept of this aspect of speech. 487 488

Discussion 489
Here we employed backward TRF models to estimate stimulus reconstruction accuracy 490 of bands of low-frequency acoustic information in the speech signal (delta-and theta-band 491 speech envelope information) along with alpha-band speech envelope information (control 492 band) in children with and without dyslexia matched for age. Two different group comparison 493 approaches (between-group and within-child) were employed to investigate whether the 494 speech-based representations for low-frequency envelope information being developed by the 495 dyslexic brain were "noisy" or "normal". If decoding accuracy varies by group for the between-496 group TRF models, then the speech-based representations of low-frequency acoustic 497 information developed by children with developmental dyslexia cannot be considered 498 "normal". However, if decoding accuracy is equal irrespective of group for within-child TRF 499 models, then the speech-based representations of low-frequency envelope information 500 developed by children with developmental dyslexia cannot be considered "noisy". The 501 modelling suggested that the dyslexic brain does not develop "noisy" representations of low-502 frequency envelope information, but neither are these representations "normal". The between-503 group models showed that the speech-based representations developed by children with 504 dyslexia were less accurate than the speech-based representations developed by children 505 without dyslexia regarding low frequency envelope information. These differences were 506 restricted in the current study to the delta band (Figure 4). The between-group modelling 507 generated lower decoding accuracies than the within-child modelling, as can be seen by 508 comparing the correlation values in Figures 4 and 6. Between-group decoding accuracy for the 509 theta and alpha bands did not exceed chance levels as estimated by the null models. Therefore, 510 the conservative conclusion is that encoding of low-frequency speech information in the delta 511 band is atypical in 9-year-old children with developmental dyslexia. 512 At the level of an individual child, however, the modelling did not support the classical 513 theoretical view (including our own earlier view, see Swan & Goswami, 1997) that the 514 phonological representations of speech developed by children with dyslexia are "fuzzy", 515 "noisy" or "imprecise" (see also Elbro et al., 1998;Snowling, 2000). Although we only 516 decoded the low-frequency envelope information in speech, the quality of the speech-based 517 representations for each child as indexed by the neural decoding method were of similar 518 consistency whether the child was dyslexic or not, for all frequency bands explored. This 519 finding suggests that even if acoustic information is weighted differently in the speech-based 520 representations developed by children with dyslexia, the experience of speech processing 521 regarding low-frequency envelope information is perceptually stable for the children 522 themselves. Nevertheless, a linear decoding approach to stimulus reconstruction is only one 523 method for assessing whether neural stimulus representations of low-frequency envelope 524 information are "noisy" or not, and converging neural methods are required. Furthermore, the 525 question of whether other features of the speech signal such as voicing may be represented in 526 a "noisy" manner by the dyslexic brain is not addressed by our data. 527 The delta-band between-group difference in decoding accuracy found here was 528 predicted by TS theory (Goswami, 2011(Goswami, , 2019, which provided the conceptual framework for 529 the current study. TS theory has proposed that dyslexic difficulties in discriminating amplitude 530 envelope rise times, difficulties found for dyslexic children in 7 languages to date (see 531 Goswami, 2015, for a review), are associated with atypical neural encoding of acoustic 532 information within the amplitude envelope < 10Hz, related to perceptual impairments in 533 processing speech rhythm. Rise times are one neural trigger for oscillatory phase-resetting, an 534 automatic process that aids multi-timescale speech-brain cortical tracking in adults (Giraud &  with dyslexia has shown that rise times do not provide an efficient phase-resetting mechanism 537 in the dyslexic brain during natural speech listening (Lizarazu et al., 2021). The between-group 538 data presented here suggest that this inefficient phase-resetting may particularly affect the 539 encoding of speech envelope information in the delta band during childhood. This delta-band 540 information is a critical feature of the speech signal. 541 The current study has several limitations. As already noted, the quality of dyslexic 542 children's speech-based representations for features other than the envelope is not addressed 543 by our backward decoding method. Further, although the backward TRF model has been used 544 as the most popular model for decoding stimulus information from neural responses, it has 545 some general limitations. Firstly, the model assumes a linear relationship between the input 546 stimuli and the neural responses. Secondly, the performance of model when generalising to 547 unseen testing data sets depends on the estimate of large number of unknown parameters and 548 on the size of the data used for training the model. It should be noted that the forward TRF 549 model also has these two latter limitations. However, many adult studies focus on backward 550 (decoding) models, as they have several advantages over forward (encoding) models, as 551 described by Crosse et al. (2016). 552 In conclusion, the current study converges with the three preceding studies of children 553 with dyslexia that measured speech-based representations directly using neural methods 554 amplitude envelope. Together with the current study, these studies suggest that the speech-557 based representations developed by children with dyslexia for low-frequency envelope 558 information <10 Hz are not "normal", supporting TS theory. Children with dyslexia encode 559 less accurate prosodic (delta band) speech-based information in their phonological 560 representations for words in comparison to typically-developing children. TS theory proposes 561 that these differences in encoding low-frequency envelope information affect phonological 562 awareness at all linguistic levels via the linguistic hierarchy, which is governed by the prosodic 563 level, and that this then impacts the process of learning to read and spell. At the same time, the 564 current study contributes a novel child-by-child neural decoding approach that reveals that the 565 perceptual world of the dyslexic child regarding low-frequency envelope information in speech 566 is a consistent one. At the level of the individual child, the speech-based representations of 567 low-frequency envelope information developed by children with dyslexia are not "fuzzy" or 568 "noisy". This finding contributes to a long-standing debate in the psycholinguistic literature 569 regarding whether the phonological 'deficit' that characterises individuals with dyslexia across 570 languages stems from differences in phonological representation or in phonological access 571 Svenkovits, 2008). The neural decoding method employed here suggests that differences in 574 phonological representation are indeed present for low-frequency speech information. This 575 conclusion accords well with the linguistic data gathered by recent longitudinal studies of 576 infants at family risk for developmental dyslexia, who show amplitude rise time discrimination 577 difficulties from age 10 months as well as poorer early word learning and poorer phonological 578 development (Kalashnikov et al., 2018, 2019a,b, 2020). Nevertheless, the strongest test of this 579 conclusion about phonological representations would be a demonstration that speech 580 production is also atypical in dyslexia regarding the specification of speech prosody. If speech 581 prosody is not encoded accurately in the dyslexic brain, yet this atypical encoding is consistent 582 and perceptually stable for the child, then speech output of prosodic information should also 583 be affected.