Acoustic Degradation Reorganizes Brain–Model Representational Geometry Across Speech and Language Hierarchies

Brain-model alignment is typically evaluated under a single stimulus condition, assuming fixed correspondence across model layers and cortical regions. We tested whether this mapping reorganizes when speech is acoustically degraded. Twenty-five participants heard 48 Mandarin sentences in Clean speech and speech-shaped noise ($-3\,\mathrm{dB}$ SNR) during fMRI. Using matched-layer representational similarity analysis ($\Delta\mathrm{RSA}$) across 14 auditory-language regions, we compared cortical patterns with a log-mel control, five speech models, and three text LLMs. Despite substantial behavioral decline, degradation produced structured, bidirectional changes: learned speech representations gained correspondence predominantly across left temporal-frontal cortex, whereas alignment with condition-invariant text-LLM geometries decreased mainly in right-hemisphere regions. Whole-search-space FDR correction confirmed increased Whisper-Tiny alignment in left temporal cortex and reduced Qwen2.5-3B alignment in right anterior temporal cortex. Acoustic degradation does not uniformly erode alignment; it selectively shifts which computational geometries best mirror cortical representations, establishing perturbational stability as an important criterion for neural alignment.