MEG recordings of listened and imagined music and speech

Magnetoencephalography (MEG) recordings of 18 participants listening to and imagining two spoken poems and two melodies, in a paired design (every participant performed both conditions for every stimulus, 10 repetitions each). Two versions are provided: Raw (raw_sub-XX.zip, 18 subjects): epoched into 27 s trials, 157 channels, 1000 Hz, no filtering or ICA. Poem tasks only. Cleaned (ica_cleaned_sub-XX.zip, 16 subjects, no sub-07 or sub-15): 155 channels, 0.1-40 Hz band-pass, per-subject ICA. Poem and melody tasks. This is the version used in the related publications. stimuli_metadata.zip contains the poem audio, transcripts, WhisperX word onsets and sensor positions. Unzip all archives into the same folder to reconstruct the directory tree. See README.md for details, loading code and citation. Related publications: M. Maghsoudi, S. S. Kankanala, S. A. Shamma, S. Ganapathy. Integrating Language Models into Listened and Imagined Speech Decoding from MEG. arXiv:2609.31997, 2026. M. Maghsoudi, S. Shamma. Zero-Shot Imagined Speech Decoding via Imagined-to-Listened MEG Mapping. arXiv:2605.08075, 2026. M. Maghsoudi, M. Rezaeizadeh, S. Shamma. A Convolutional Framework for Mapping Imagined Auditory MEG into Listened Brain Responses. 2025 59th Asilomar Conference on Signals, Systems, and Computers, pp. 1582-1586.