Search Abstracts | Symposia | Slide Sessions | Poster Sessions
How do neural dynamics of the linguistic hierarchy scale with the speed of speech?
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Xirong Hu1, Yi Li2, Alexandra Woolgar2, Laura Gwilliams1; 1Stanford University, 2MRC Cognition and Brain Sciences Unit, University of Cambridge
Speech comprehension requires building a linguistic hierarchy of neural representations over time. These features have characteristic neural timescales, with low-level features unfolding faster and high-level features unfolding more slowly (Gwilliams et al., 2025). Classic language studies have long associated specific neural processes with specific response latencies: for example, M100 encodes phonetic information at 100ms, and the N400 responds to semantic anomalies at 400 ms. Here we ask a simple but fundamental question: Do the neural timescales of speech responses at different hierarchical levels scale with the speed of the input (rate-adapted hypothesis), or remain fixed (rate-invariant hypothesis)? We recorded MEG from 23 native English speakers (aged 21–34) while they listened to natural speech (e.g., audiobooks) presented at three speeds (1x, 1.5x, 2x). Audiobooks were annotated for phonetic and lexical features. We regressed out speech envelope and pitch contribution from the MEG signal using a TRF model, and then applied linear decoding to this residual MEG signal using time-resolved logistic regression classification. At the phoneme level, we decoded 14 binary acoustic-phonetic features (covering manner, place, and voicing). At the word level, we decoded 4 features: noun, verb, adjective and lexical frequency (high vs. low). We ran two decoding analyses. First, we generate a single time-course of decoding performance (AUC) for each feature and speed to test performance and duration of decoding at different input speeds. Second, we test the generalisation of decoding across different speech rates using the temporal generalisation approach (King and Deheane, 2014). This allows us to draw direct comparisons between the neural encoding of speech features at different speeds. The majority of phonetic and lexical features were decodable from MEG signals at all speeds (p<0.01). Phonetic features localised to sensors over auditory cortex, with increased right lateralisation for vowels; lexical features localised to frontal sensors. Phonetic features and lexical frequency showed decreased decoding performance as speech rate increased, whereas part of speech features did not show this decrease. Overall decoding duration decreased at faster speeds for all features, and the relationship was non-linear (e.g., speech that was 2x faster had a ~70% duration). For temporal generalisation, training on one speed and testing on another speed, we observed above-chance decoding performance, suggesting that the neural processes are common across speeds. Finally, when evaluating the timing of these neural processes, we found that responses occurred earlier for faster speeds, but later than that would be predicted from purely linear compression. Our results are inconsistent with both rate-invariant hypothesis and rate-adapted hypothesis; instead, they support a “goldilocks” hypothesis that sits somewhere in-between. Neural processing speeds up when the speech is faster, but does not show full rate adaptation. Furthermore, this adaptation differs across features in the linguistic hierarchy: phoneme and lexical frequency representations are compressed and attenuated at faster rates, while part-of-speech representations are stable in performance across speeds. This dissociation might indicate that different levels of the speech-processing hierarchy adapt to accelerated speech through different mechanisms, with high-level lexical representations preserved despite degraded bottom-up signals.
Topic Areas: Speech Perception, Computational Approaches