Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Hearing speech across voices: EEG evidence for the role of speaker variability in statistical learning of word boundaries
Poster Session F, Friday, October 2, 2:45 - 4:45 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Tianze Xu1, Meng Yao1, Alexis Hervais-Adelman2,1, Nathalie Giroud1, Volker Dellwo1, Stefan Elmer1; 1University of Zurich, 2University of Geneva
Comprehending spoken language requires segmenting a continuous acoustic signal into linguistic units such as words. A crucial mechanism in developing the discrete targets for speech segmentation is statistical learning, which arises through extraction of transitional probabilities between adjacent syllables in continuous speech. Although neural correlates of statistical learning have been extensively investigated using electroencephalography (EEG), most studies have employed single-speaker input, leaving unresolved how variability in speaker identity influences this process. This question is theoretically important because natural speech perception requires listeners to process words under conditions of constant speaker variation. The planned study investigates how speaker variability modulates statistical learning of word boundaries and the neural dynamics supporting this process. Adult native speakers of German (target N = 50) will listen to continuous streams of consonant–vowel syllables constructed following German phonotactics. The design crosses speaker consistency (single-speaker baseline, boundary-aligned speaker change, boundary-misaligned speaker change) with statistical structure (structured vs. pseudo-random streams) in a within-subjects 3 × 2 design. In structured streams, four trisyllabic (pseudo-)words are repeated continuously, producing high transitional probabilities between adjacent syllables within words (1.00) and lower probabilities across word boundaries (0.33). In pseudo-random streams, the same syllables are shuffled such that transitional probabilities are uniformly low (~0.09). Two speaker identities will be created through controlled morphing of glottal pulse rate (fundamental frequency) and vocal tract length (formant dispersion). In the boundary-aligned condition, changes between the two speaker identities always coincide with word boundaries, whereas in the boundary-misaligned condition, speaker changes occur pseudo-randomly from syllable to syllable and do not align with word boundaries. Each exposure stream lasts approximately nine minutes and is divided into three blocks. After each block, participants complete a four-trial two-alternative forced-choice task assessing word learning by selecting between a word encountered during exposure and a part-word (i.e., a trisyllabic sequence spanning word boundaries). EEG will be recorded continuously during exposure and analysed with a focus on two complementary neural indices of statistical learning. First, event-related potentials (ERPs) will assess the emergence of proto-lexical representations, particularly the frontocentral N400 typically associated with word learning in structured streams. Second, frequency-domain analyses will quantify neural entrainment at word frequency using inter-trial coherence (ITC) and spectral power measures. Several theoretically informative outcomes are possible. If speaker changes aligned with word boundaries facilitate learning, enhanced behavioural performance, stronger word-frequency neural entrainment (indexed by ITC), and increased frontocentral N400 responses may emerge in the boundary-aligned speaker change condition relative to the single-speaker baseline. In contrast, boundary-misaligned speaker change may disrupt statistical learning by introducing competing cues, resulting in lower behavioural performance, reduced word-frequency ITC, and attenuated N400 effects. Alternatively, if listeners successfully normalise across speaker variability, statistical learning and its neural signatures may remain robust across all three speaker consistency conditions. The study extends statistical learning research into a more ecologically realistic domain by incorporating speaker variability and provides one of the first neuroimaging investigations of how speaker identity interacts with speech segmentation mechanisms.
Topic Areas: Speech Perception, Language Development/Acquisition