Artificial intelligence for brain-to-speech decoding in paralysis: a systematic review

The loss of communication constitutes a critical challenge for people living with paralysis. Brain-computer interfaces (BCIs) paired with artificial intelligence (AI) provide an opportunity to restore this ability. This systematic review examined the use of AI to decode speech from brain signals through both invasive and non-invasive neural interfaces. Using the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines, we reviewed 115 studies published between 2019 and 2025 to extract data on acquisition protocols, signal preprocessing, and AI architectures. The quality of each study was evaluated using Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2). The results indicated that the invasive approach achieved a higher median multiclass classification accuracy than the non-invasive method (77.7%, interquartile range (IQR): 61.4-93.2% vs. 73.0%, IQR: 49.3-89.2%; N = 82 studies with comparable multiclass metrics), computed from the best-performing model per study across heterogeneous task types, vocabulary sizes, and predominantly subject-dependent evaluation paradigms (84.3% of studies). However, this narrow gap in raw accuracy (4.7% points) should not be interpreted as a direct cross-modality performance ranking, as it obscures substantial differences in task complexity (invasive studies typically decoded larger vocabularies and continuous speech), evaluation paradigm, and participant population. Additionally, the hybrid convolutional neural network/recurrent neural network (CNN/RNN) architecture and transformers outperformed traditional classifier models. Nevertheless, the quality assessments showed significant limitations; notably, 62.6% of the studies evaluated had a high risk of selection bias due to patient characteristics, and only six studies (5.2%) validated results in paralyzed individuals-all relying on invasive modalities. Among these, classification accuracy ranged from 47.1% to 90.0%, while word error rates for continuous speech decoding ranged from 25.6% to 58.8%, demonstrating feasibility but with substantial variability across paralyzed cohorts. No non-invasive study has demonstrated functional speech decoding in paralyzed populations. This validation gap represents the most urgent translational priority in this field. We proposed a decision framework to address accuracy, cost constraints, and clinical applicability.