Can Tokens Compete? Token Representations against Supervised CNN Backbones for BirdCLEF+ 2026

arXiv cs.AI · 6 hours ago

A research team developed a supervised baseline for BirdCLEF+ 2026 bird vocalization detection that combines Perch v2, HGNetV2-B0, and a prototypical head, achieving a leaderboard score of 0.936. They then evaluated whether token-based representations from neural audio codecs and foundational embeddings could match or exceed the performance of bioacoustic specialist models on this multi-label animal vocalization task. The work explores the viability of token representations as alternatives to traditional supervised CNN backbones for wildlife sound classification.