Reproduce “On Batching Variable Size Inputs for Training End-to-End Speech Enhancement Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Philippe Gonzalez, Tommy Sonne Alstrøm, Tobias May
Researchers at the Technical University of Denmark demonstrated that sorted or bucket batching, when combined with dynamic batch sizes, substantially reduces training time and GPU memory usage for end-to-end speech enhancement systems, achieving comparable or better generalization performance than traditional random batching.
Key insights
- Strict within-batch randomization is not essential for robust generalization in speech enhancement, as sorted batching performed comparably to random batching.
- Smaller batch sizes consistently lead to improved generalization performance, particularly under challenging mismatched acoustic conditions.
- Dynamic batch sizing effectively optimizes GPU memory allocation by maintaining a more consistent total data amount per batch for variable-length inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.