A Novel Machine Learning Model for Non-Invasive EEG-Based Inner-Speech Translation in ALS

Amyotrophic lateral sclerosis (ALS) progressively paralyzes speech and motor functions, rendering patients speechless when voluntary movement is lost. Although invasive brain–computer interfaces (BCIs) such as electrocorticography (ECoG) have achieved high accuracy in speech-related neural activity decoding, clinical use is limited by surgery costs and risks. This study presents a new, fully non-invasive EEG-based inner-speech translation system aimed at communication restoration in ALS. This paper introduces a novel machine learning architecture based on a fusion of region-aware convolutional attention encoders and a transformer decoder, presenting a generative sequence-to-sequence model for open-vocabulary EEG-to-text translation. Preprocessed and downsampled EEG data from the Chisco imagined-speech dataset were used, and 48 optimized electrode channels were selected based on contribution-based ranking. Parameter-efficient fine-tuning (LoRA, PEFT) was employed to train the model to preserve linguistic fluency while being specialized for EEG input. Evaluation achieved BLEU-1 = 0.512 $\pm$ 0.012 and ROUGE-L F1 = 0.396 $\pm$ 0.009, surpassing earlier non-invasive baselines. Ablation studies confirmed the critical role of cross-regional fusion and diversity regularization in maintaining cortical interpretability during the encoding process. To the best of our knowledge, this is the first open-vocabulary, non-invasive EEG-to-text system capable of reconstructing continuous inner speech with high linguistic coherence and accuracy. These findings are a key step toward clinical neural speech restoration for ALS, facilitating the development of future real-time, low-cost assistive communication devices.