English

Input Conditioned Layer Dropping in Speech Foundation Models

Sound 2025-07-11 v1 Computer Vision and Pattern Recognition Audio and Speech Processing

Abstract

Curating foundation speech models for edge and IoT settings, where computational resources vary over time, requires dynamic architectures featuring adaptable reduction strategies. One emerging approach is layer dropping (LD\mathcal{LD}) which skips fraction of the layers of a backbone network during inference to reduce the computational load. This allows transforming static models into dynamic ones. However, existing approaches exhibit limitations either in the mode of selecting layers or by significantly modifying the neural architecture. To this end, we propose input-driven LD\mathcal{LD} that employs the network's input features and a lightweight layer selecting network to determine the optimum combination of processing layers. Extensive experimentation on 4 speech and audio public benchmarks, using two different pre-trained foundation models, demonstrates the effectiveness of our approach, thoroughly outperforming random dropping and producing on-par (or better) results to early exit.

Keywords

Cite

@article{arxiv.2507.07954,
  title  = {Input Conditioned Layer Dropping in Speech Foundation Models},
  author = {Abdul Hannan and Daniele Falavigna and Alessio Brutti},
  journal= {arXiv preprint arXiv:2507.07954},
  year   = {2025}
}

Comments

Accepted at IEEE MLSP 2025

R2 v1 2026-07-01T03:55:10.586Z