SOUND#
SOUND estimates channel-specific noise using a lead field and iteratively constructs a linear sensor-space cleaning operator [1][2].
Usage#
from mne_denoise.sound import SOUND
model = SOUND(reference="best")
clean = model.fit_transform(raw)
Pass compatible EEG MNE Raw, Epochs, or Evoked objects with a montage directly;
without forward, they can use the spherical EEG lead-field fallback. MEG or
mixed-channel MNE input requires an explicit forward solution. NumPy input also
requires an explicit lead field with matching channel order and the corresponding
array layout documented by the API.
Key points#
reference=”best” selects a low-noise single-channel reference for the estimation and maps the result back to an average-referenced full-channel operator. reference=”average” uses the all-channel solver and assumes average-referenced input.
forward supplies an individual lead field. Compatible EEG MNE input with a montage can use the spherical fallback; MEG or mixed-channel MNE input and NumPy input require an explicit forward.
lambda_, n_iter, and tol control regularization and convergence. sigma_source=”evoked” estimates epoched noise from the trial average; “trials” uses concatenated trials.
The fitted operator_, sigmas_, convergence_, and best_channel_ are useful diagnostics. The selected reference index is returned by compute_sound_ref_best.
Re-referencing the data or lead field changes the fitted operation. Use the same channel order and reference for both.
SOUND reduces estimated sensor noise; evaluate any change to the signal of interest separately.