Encoder layers in speech enhancement models maintain noise-invariant representations while decoder layers adapt strongly to degradation, with the pattern consistent across architectures and degradation types.
Model and Activation Map The model chosen for the analysis is MUSE [3], a transformer- convolutional speech enhancement model trained on the V oiceBank-DEMAND dataset [4, 5]
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
eess.AS 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Where Does Speech Enhancement Adapt? Probing Study Under Controlled Degradation
Encoder layers in speech enhancement models maintain noise-invariant representations while decoder layers adapt strongly to degradation, with the pattern consistent across architectures and degradation types.