zero-shot-imagined-speech-meg

name

zero-shot-imagined-speech-meg

description

Zero-shot imagined speech decoding from MEG via imagined-to-listened cross-condition mapping. Trains models to map imagined MEG responses to listened responses, then decodes using listened-only decoder. Three-stage pipeline: (1) mapping imagined→listened MEG, (2) train contrastive word decoder on listened MEG with multi-embedding evaluation, (3) decode imagined speech via mapping pipeline on held-out subjects. Use when: imagined speech decoding, MEG BCI, cross-condition neural mapping, zero-shot brain decoding, non-invasive speech BCI. Activation: imagined speech, MEG decoding, brain-computer interface, speech neuroprosthetics

Zero-Shot Imagined Speech Decoding via Imagined-to-Listened MEG Mapping

Three-stage pipeline that maps imagined MEG to listened MEG representations, enabling zero-shot imagined speech decoding using listened-only training data.

Metadata

• Source: arXiv:2605.08075
• Authors: Maryam Maghsoudi, Shihab Shamma
• Published: 2026-05-08
• Subjects: cs.LG; eess.AS

Core Methodology

Key Innovation

Imagined speech datasets are scarce and hard to align temporally. This approach circumvents the data scarcity problem by:

1. Collecting paired listened + imagined MEG from trained musicians (improved temporal alignment)
2. Training models to map imagined → listened MEG representations
3. Using the mapping to enable imagined speech decoding via a listened-only decoder

Three-Stage Pipeline

Stage 1: Imagined-to-Listened Mapping

• Train 6 linear and neural models mapping imagined MEG → listened MEG
• Validate against null baseline from unseen subjects
• Verify predicted-listening responses preserve stimulus-specific information