Imagined Speech Reconstruction with 3D Neural Metabolism and Large Language Model Integration

Abstract

Cognitive linguistics posits that language underpins human thought, and this principle has influenced the study and development of large language models (LLMs). In particular, several studies have investigated the metabolic costs of sentence formation using neuroimaging techniques such as positron emission tomography, functional magnetic resonance imaging, electroencephalography (EEG), and imagined speech reconstruction (ISR). In this study, EEG data corresponding to imagined English-language speech phonemes were used for ISR, in combination with an LLM trained on an abridged autobiography. The LLM-generated text responses guided the synthesis of EEG data from relevant phonemes, which were then used to estimate corresponding metabolic activity, and the changes in simulated neurometabolic and electrical parameters were visually represented. Notably, introducing pseudorandom variance significantly (p < 0.001) enhanced the models ability to reflect biological variability. Future directions include expanding the ISR system with lightweight or locally run LLMs, incorporating training data from larger and more diverse populations, and utilizing truly random variability sources. Further optimization for broader hardware compatibility and implementation--such as neural phantoms, emotional context integration, or human-computer interaction platforms--offer promising pathways for advancement. Overall, this work establishes a foundation for the next generation of biologically inspired, modular, and adaptable ISR systems for both research and practical applications.