Fig. 5: Bilingual listeners show neural encoding of word-level features and phoneme surprisal for both familiar languages.

a, Speech-responsive electrodes across eight Spanish–English bilingual participants show similar STG localization for Spanish (left) and English (right) speech across hemispheres. Electrode size indicates magnitude of the peak HFA averaged sentence response. Inset histograms show electrode counts across English and Spanish speech in each anatomical region. b, TRF weights across speech-responsive electrodes for English (top) and Spanish (bottom) are significantly correlated (Pearson r(804) = 0.88, P < 0.001). Top, columns (electrodes) ordered by the acoustic–phonetic feature showing the maximal weight. Bottom, columns are ordered identical to top for visual comparison. c, Distribution of TRF Spanish–English weight correlations across electrodes (purple) is significantly higher compared to the non-parametric permuted distribution (grey). Black vertical line indicates the 95th percentile in the permuted distribution. d, Encoding of word-level and phoneme-surprisal features is not significantly different across Spanish and English (LME speech condition P > 0.05; see Methods for formula). e, Encoding of word-level and phoneme-surprisal features is not significantly different between bilingual and the native language for monolingual participants (LME speech condition P > 0.05; see Methods for formula). f, Example electrode from a Spanish–English bilingual participant differentiates word and syllable boundaries in both the English speech condition (left) and Spanish speech (right). g, Forty-five electrodes show positive unique variance for word-level and phoneme surprisal in English and Spanish. Unique variance for word-level features across languages is significantly correlated (Spearman r(45) = 0.31, P = 0.03). Inset plot shows the count of electrodes in each quadrant of the scatter plot. h, World map indicating language backgrounds represented in the participant cohort and extent to which these languages are spoken across the world. Maps generated using MathWorks mapping toolbox and information from the World Factbook¹⁰⁵. i, Neural word-boundary classification AUC in English scales with English language proficiency for bilingual speakers of Spanish–English and speakers of diverse languages (LME proficiency P < 0.001; see Methods for formula). Each line corresponds to the mean AUC per participant; each scatter point is AUC of 1/15 folds.