MindAlign: Decoding Inner Speech from fMRI Signals via Multimodal Embedding Alignment under Limited Data

MindAlign introduces a two-stage brain-to-language framework that decodes inner speech from fMRI signals, addressing challenges such as limited training data and inter-subject variability. The first stage establishes a subject-specific neural-semantic alignment, while the second stage uses this alignment to prompt a frozen multimodal language model for text generation. Experimental results demonstrate superior performance over existing methods, highlighting the framework's potential for scalable and adaptable brain-to-text applications in AI.