Brain2Text: A Reproducible, CPU-Deployable Framework for Neural Speech Decoding with Browser-Accessible Inference

Despite the highly successful results neural speech decoding models have obtained in laboratories, the road to making these models usable by clinicians and end-users is under reported and often not shown. The published decoders are embedded within the jupyter notebooks, are displayed in a terminal and require GPUs on research infrastructure to operate. Brain2Text was developed to fill right up this hole. The framework receives pre-extracted intracortical feature vectors of dimension 512 from the Brain-to-Text 2025 T15 CopyTask benchmark, and outputs English words on their basis using the following five steps: (1) 512 dimensional feature vectors are extracted from the intracortical areas within the benchmark, (2) a five-layer Gated Recurrent Unit (GRU) network is trained with the extracted feature vectors and a Connectionist Temporal Classification (CTC) loss function, (3) feature vectors are mapped to the output language (English) using a frequency-weighted CMU Pronouncing Dictionary ( lookup) , (4) an LLM fallback on the unmapped phoneme sequences. The end-to-end inference pipeline runs on CPU with latency of 90-165 ms for trials of up to 200 time-steps. The decoder is wrapped in a Flask REST API which is accessed by a React/Vite front end application, or used in a live-less demonstration mode in which there are no dependencies on a live back end or files. The whole stack is initialised by one command in the shell.