Speech language models lack important brain-relevant semantics

TL;DR

The study probes what brain alignment with language models actually reflects by directly removing low-level textual, speech, and visual features from model representations and measuring changes in fMRI alignment during reading and listening to the same naturalistic narratives. It compares text-based models (BERT, GPT-2, FLAN-T5) with speech-based models (Wav2Vec2.0, Whisper) using a ridge-removal residual approach and voxelwise encoding with cross-subject normalization. Key findings show that late-language-region alignment for text models persists despite feature removal, indicating brain-relevant semantics beyond low-level cues, while early-sensory alignments are largely driven by textual features; for speech models, early auditory alignment remains strong after removal but late-region alignment collapses, suggesting limited semantics in higher-order regions. These results highlight the need for cautious interpretation of brain–model similarity and suggest integrating strengths of both model families to better capture the full spectrum of brain-language processing.

Abstract

Despite known differences between reading and listening in the brain, recent work has shown that text-based language models predict both text-evoked and speech-evoked brain activity to an impressive degree. This poses the question of what types of information language models truly predict in the brain. We investigate this question via a direct approach, in which we systematically remove specific low-level stimulus features (textual, speech, and visual) from language model representations to assess their impact on alignment with fMRI brain recordings during reading and listening. Comparing these findings with speech-based language models reveals starkly different effects of low-level features on brain alignment. While text-based models show reduced alignment in early sensory regions post-removal, they retain significant predictive power in late language regions. In contrast, speech-based models maintain strong alignment in early auditory regions even after feature removal but lose all predictive power in late language regions. These results suggest that speech-based models provide insights into additional information processed by early auditory regions, but caution is needed when using them to model processing in late language regions. We make our code publicly available. [https://github.com/subbareddy248/speech-llm-brain]

Figures (20)

• Figure 1: A direct approach to test the effect of low-level stimulus features on the alignment between different types of language models and brain recordings (reading vs. listening).
• Figure 2: Contrast of estimated cross-subject prediction accuracy for reading and listening for a representative subject (subject-8). Blue and Red voxels depict higher cross-subject prediction accuracy estimates during listening and reading, respectively. Voxels that have similar cross-subject prediction accuracy during reading and listening appear white, and are distributed across language regions. Here, middle frontal gyrus (MFG), inferior frontal gyrus (IFG), inferior frontal gyrus orbital (IFGOrb), angular gyrus (AG), and lateral temporal cortex (LTC) are late language regions, EVC denotes early visual cortex and AC denotes auditory cortex. Cross-subject prediction accuracy for other participants are reported in Appendix Figs. \ref{['fig:noise_ceiling_subject01']} and \ref{['fig:noise_ceiling_subject07']}.
• Figure 3: ROI-based normalized brain alignment was computed by averaging across participants, models, layers, and voxels. Orange: text, Green: speech, Solid: reading, Patterned: listening. Red dashed line: chance performance, and * at a particular bar indicates that the model's prediction performance is significantly better than chance.
• Figure 4: Reading condition in the Early Visual and Late Language Regions: (a) For Text-based Models: Average normalized brain alignment was computed across participants before and after removal of low-level stimulus features, across layers and voxels. (b) For Speech-based Models: Similar alignment analysis was conducted for speech models before and after removal of low-level stimulus features, across layers and voxels. Red dashed line: chance performance, and * indicates that the residuals prediction performance is significantly better than chance.
• Figure 5: Listening condition in the Early Auditory and Late Language Regions: (a) For Text-based Models: Average normalized brain alignment was computed across participants before and after removal of low-level stimulus features, across layers and voxels. (b) For Speech-based Models: Similar alignment analysis was conducted for speech models before and after removal of low-level stimulus features, across layers and voxels. Red dashed line: chance performance, and * indicates that the residuals prediction performance is significantly better than chance.