Proceedings of the Annual Meeting of the Cognitive Science Society

Direct speech synthesis from neural activity offers a lifeline for severe paralysis, yet decoding high-dimensional, noisy signals remains a formidable challenge. Current approaches rely on passive regression, overlooking the hierarchical and generative nature of speech production. We propose Neuro-Composer, a biologically grounded framework reframing decoding as the dynamic integration of top-down semantic planning and bottom-up acoustic execution. Grounded in Dual-Stream Theory, our architecture decomposes decoding into a ventral stream extracting abstract intent via foundation models (BERT/Wav2Vec 2.0), and a dorsal stream capturing fine-grained articulatory dynamics. These streams are synthesized via a latency-aware Predictive Refiner. A bio-mimetic auditory feedback loop enforces semantic consistency, endowing the system with self-monitoring capabilities. Experiments demonstrate Neuro-Composer significantly outperforms baselines in reconstruction fidelity. Analyses reveal emergent semantic manifolds and brain-like gating patterns, providing computational validation for neurocognitive theories of language production.