From Artificial Neurons to Speech Neuroprostheses: A Review of Neural Networks, NeuroAI, and Brain–Computer Interfaces for Communication Without Vocalization

Artificial neural networks were originally inspired by the brain, and they are now returning to neuroscience both as quantitative models of cortical computation and as the decoding engines of brain–computer interfaces. This review traces that two-way relationship and evaluates how close current technology is to letting people communicate by thought alone, without moving the vocal tract. It contrasts biological and artificial neurons, summarizes the development of deep learning from threshold units to transformers, and reviews evidence that deep networks predict activity in visual and language cortex, along with advances in connectomics and neuromorphic hardware. The core of the review analyzes speech neuroprostheses reported from 2019 to 2026. Implanted systems that decode attempted speech from motor cortex now reach word accuracies above 97% with vocabularies of more than 125,000 words, at roughly 30 to 80 words per minute, and one participant has used such a system independently at home for nearly two years. Decoding of purely inner speech has been demonstrated with lower accuracy and raises mental-privacy concerns that can be mitigated by design. Noninvasive decoding remains far from practical communication. Industrial programs, ethics, and emerging regulation of neural data are also examined.This is a preprint and has not been peer reviewed. AI assistance (Claude, Anthropic) was used in preparing this manuscript, as disclosed in the Acknowledgment section.