Fig. 3: Enhanced discrimination between word and syllable boundaries in native speech.

a, Spectrogram and speech envelope for example Spanish (top) and English (bottom) sentence stimuli. Word and syllable boundaries marked in solid and dashed lines, respectively. b, Spectrogram and speech envelope for example word and syllable boundaries are closely matched to one another in Spanish (top) and English (bottom) but differ across languages. Example word and syllable boundaries marked as in a. c, Acoustic classification of word boundaries using the spectrogram across English and Spanish speech. Each scatter point represents 1/20 folds. Permuted distribution median AUC of 0.50 (25 permutations with 20 folds each for each language). Box plots show maximum and minimum values (whiskers), median (centre line) and 25th to 75th percentiles (box limits). d, Left, example electrode (E1) from an English listener differentiates word and syllable boundaries in native (top) but not in foreign (bottom) speech. Rectangular patches indicate time points with a significant difference (per-time-point two-sided ANOVA P < 0.01 with Bonferroni correction). Continuous shaded patches indicate standard error of the mean HFA response across boundaries. Right, anatomical location of E1 and the unique variance for word and phoneme-surprisal features in the example participant. e, Left, example electrode (E2) from a Spanish listener differentiates word and syllable boundaries in native (top) but not in foreign (bottom) speech. Rectangular patches indicate time points with a significant difference (per-time-point two-sided ANOVA P < 0.01 with Bonferroni correction). Continuous shaded patches indicate standard error of the mean HFA response across boundaries. Right, anatomical location of E2 and unique variance for word and phoneme-surprisal features in the example participant. f, Word versus syllable boundary discrimination is significant in more electrodes (left) and for a longer duration of the neural response (right) in native versus foreign speech (LME ***P < 0.001; see Methods for formula).