mne_denoise.dss.denoisers.SpectrogramDenoiser#

class mne_denoise.dss.denoisers.SpectrogramDenoiser(threshold_percentile: float = 90.0, nperseg: int = 256, noverlap: int | None = None, mask: ndarray | None = None)[source]#

Adaptive/Nonlinear spectrogram denoiser (Section 4.1.3).

Applies masking in the time-frequency domain. This version is ADAPTIVE, calculating the mask from the source estimate itself at each iteration. This makes it distinct from the Linear SpectrogramBias.

Parameters:

  • threshold_percentile (float) – For adaptive masking, threshold below this percentile. Default 90. Higher percentile = sparser signal (more aggressive denoising).

  • nperseg (int) – Segment length for STFT. Default 256.

  • noverlap (int, optional) – Overlap between segments. Default nperseg // 2.

  • mask (ndarray, shape (n_freqs, n_times), optional) – Optional FIXED mask to use instead of adaptive (hybrid mode).

Examples

>>> from mne_denoise.dss.denoisers import SpectrogramDenoiser
>>> # Retain only the strongest 10% of TF-bins (aggressive denoising)
>>> denoiser = SpectrogramDenoiser(threshold_percentile=90)
>>> denoised = denoiser.denoise(source)

See also

SpectrogramBias

Fixed linear version.

References

Särelä & Valpola (2005). Section 4.1.3 “SPECTROGRAM DENOISING”

__init__(threshold_percentile: float = 90.0, nperseg: int = 256, noverlap: int | None = None, mask: ndarray | None = None) → None[source]#

Methods

__init__([threshold_percentile, nperseg, ...])

denoise(source)

Apply adaptive 2D spectrogram masking.