A Semantically Consistent Dataset for Data-Efficient Query-Based Universal Sound Separation скачать в хорошем качестве

A Semantically Consistent Dataset for Data-Efficient Query-Based Universal Sound Separation

Researchers addressing the persistent issue of residual noise in universal sound separation systems identify that current performance limitations stem largely from co-occurrence bias in standard training datasets, where models mistakenly learn to associate background noise with target sounds. To overcome this data bottleneck, the authors introduce an automated pipeline that utilizes large multimodal models to mine high-purity, single-event audio segments from wild data, resulting in a new high-quality synthetic dataset called Hive. By rigorously filtering for semantic consistency and employing a logical mixing strategy that prevents unnatural sound combinations, this approach prioritizes the quality of supervision signals over mere quantity. Remarkably, models trained on Hive's 2,400 hours of curated audio achieved separation accuracy and perceptual quality competitive with state-of-the-art foundation models like SAM-Audio, which was trained on a dataset nearly 500 times larger, thereby proving that optimizing data purity offers a highly efficient alternative to brute-force scaling in developing robust auditory AI. https://arxiv.org/pdf/2601.22599