Reproduce “You Only Hear Once: A YOLO-like Algorithm for Audio Segmentation and Sound Event Detection” 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
Satvik Venkatesh, David Moffat, Eduardo Reck Miranda
The Interdisciplinary Centre for Computer Music Research (ICCMR) at the University of Plymouth developed "You Only Hear Once (YOHO)," an algorithm for audio segmentation and sound event detection that redefines the task as a direct regression problem. This approach, inspired by YOLO from computer vision, significantly accelerates inference and post-processing times by 6-14x and 7x respectively, while achieving competitive or superior accuracy on various music-speech and environmental sound datasets.
Key insights
- Audio segmentation and sound event detection can be effectively reframed as a regression problem for predicting event boundaries, rather than a frame-based classification task.
- Significantly reducing the number of output predictions and simplifying post-processing leads to substantial improvements in inference speed without compromising accuracy.
- The YOHO paradigm, drawing inspiration from computer vision's YOLO, demonstrates that end-to-end learning for temporal audio events is feasible and highly efficient.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.