Neural encoding of the speech envelope by children with developmental dyslexia

Abstract

Developmental dyslexia is consistently associated with difficulties in processing phonology (linguistic sound structure) across languages. One view is that dyslexia is characterised by a cognitive impairment in the "phonological representation" of word forms, which arises long before the child presents with a reading problem. Here we investigate a possible neural basis for developmental phonological impairments. We assess the neural quality of speech encoding in children with dyslexia by measuring the accuracy of low-frequency speech envelope encoding using EEG. We tested children with dyslexia and chronological age-matched (CA) and reading-level matched (RL) younger children. Participants listened to semantically-unpredictable sentences in a word report task. The sentences were noise-vocoded to increase reliance on envelope cues. Envelope reconstruction for envelopes between 0 and 10Hz showed that the children with dyslexia had significantly poorer speech encoding in the 0-2Hz band compared to both CA and RL controls. These data suggest that impaired neural encoding of low frequency speech envelopes, related to speech prosody, may underpin the phonological deficit that causes dyslexia across languages.

Keywords: Dyslexia; Oscillations; Phonology; Rhythm.

Fig. 1

The noise-vocoding technique. The speech is filtered into a number of frequency bands. Within each frequency band the envelope is obtained by way of the Hilbert transform. This envelope is then used to modulate band pass filtered noise of the same bandwidth as the initial frequency band. These speech-amplitude modulated narrow-band noise signals are then recombined resulting in the noise-vocoded speech signal. Noise-vocoding degrades the spectral content of the speech. However, since the envelope is maintained throughout the noise-vocoding process the overall envelopes before and after vocoding are preserved. The number of frequency bands can be chosen at will. The lower the number of frequency bands used, the more degraded and unintelligible the speech is. Here we employed 8-channel vocoding to ensure some intelligibility while keeping performance from reaching ceiling.

Fig. 2

Stimulus reconstruction accuracy by group. Panel A shows the accuracy of reconstruction as assessed by the Pearson correlation between the actual stimulus envelope and the EEG reconstruction in each of the tested frequency bands. Black bars indicate the 95% confidence interval. Panel B shows group average envelope reconstructions in the 0–2 Hz band for a representative stimulus (Roles teased his drain) (coloured traces). The shaded coloured area shows the 95% confidence intervals of the reconstructions. The black trace indicates the actual 0–2 Hz envelope of the stimulus. Panel C shows the average frequency spectrum of the envelopes of the stimuli. Indicated are the modal frequency (1.69 Hz, which coincides with the syllable rate) and the prosody rate (0.92 Hz).

Fig. 3

Partial correlation between phonological awareness and mean reconstruction accuracy, controlling for age and IQ. The figure shows the relationship (Spearman’s ρ) between mean reconstruction accuracy and performance on the lexical stress perception task. The scatter plot shows residual variables after removing the variability due to age and IQ using linear regression. The reconstruction accuracy measure is from the overall analysis, where performance on the word report task is not matched.

Fig. 4

Group topography effects of beta (18–22Hz) power. Significant beta power differences at the topographical level between groups. Clusters that reached the significance threshold and survived correction for multiple comparisons are highlighted in blue (P_{cluster-corrected} < 0.05, based on non-parametric permutation analysis and cluster-based correction for multiple comparisons using 1000 random data partitions).