Reproduce “Audio Enhancement for Computer Audition -- An Iterative Training Paradigm Using Sample Importance” 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
Manuel Milling, Shuo Liu, Andreas Triantafyllopoulos, Ilhan Aslan, Björn W. Schuller
Researchers from the University of Augsburg, Technical University of Munich, Imperial College London, and Huawei Technologies developed an iterative training paradigm that uses sample importance to jointly optimize audio enhancement and computer audition models. This approach consistently outperformed traditional baselines and state-of-the-art denoising, recovering up to 89% of performance loss at 0 dB SNR in tasks like speech emotion recognition by enabling task-specific audio enhancement.
Key insights
- Jointly optimizing audio enhancement (AE) and computer audition task (CAT) models results in superior performance compared to independent training or data augmentation.
- Leveraging sample importance, derived from the CAT's loss, effectively directs the AE module to focus its enhancement efforts on challenging samples, particularly in low Signal-to-Noise Ratio conditions.
- The audio enhancement module can adapt its strategy to be task-specific, performing both speech enhancement (e.g., for ASR) and speech suppression (e.g., for Acoustic Scene Classification).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.