The Semantic Bottleneck: Leveraging Semantic Representations for Non-Invasive Speech Decoding
The paper argues that MEG speech decoding works better when it targets sentence meaning before text.
Brain2Semantics2Text maps sentence-level MEG responses into a semantic embedding space, then converts those predicted embeddings back into natural language. The authors frame this as a way around the low signal-to-noise limits that make phoneme- or word-level reconstruction hard in non-invasive recordings. They report improved sentence-level results compared with prior non-invasive Brain2Text methods. Source: HF Daily Papers' note
Brain2Semantics2Text maps sentence-level MEG responses into a semantic embedding space, then converts those predicted embeddings back into natural language. The authors frame this as a way around the low signal-to-noise limits that make phoneme- or word-level reconstruction hard in non-invasive recordings. They report improved sentence-level results compared with prior non-invasive Brain2Text methods. Source: HF Daily Papers' note
score 5