Official repo for the paper Input Conditioned Layer Dropping in Speech Foundation Models accepted at IEEE Machine Learning and Signal Processing (MLSP) 2025.
The Paper is available arxiv OpenReview.
This paper targets to convert a static foundation model into a dynamic foundation model, by employing "Layer Dropping" technique to skip / drop sub-modules of the encoder. We proposed a lightweight plug-and-play "Layer Selecting" block that allows skipping / dropping of encoder sub-modules based on the current input.
More Details coming soon. Stay Tuned !!