Automated Bird Species Detection From Soundscapes Using Deep Embeddings and State-Space Models

Automated biodiversity monitoring from passive acoustic recordings is a critical yet technically demanding tool for conservation science. This paper addresses the challenge of identifying 234 target fauna classes within continuous 60-second soundscape recordings from the Brazilian Pantanal wetlands. A hierarchical multi-stage pipeline is developed using Google's pretrained Perch audio model, per-species MLP probes, a custom bidirectional Selective State-Space Model (LightProtoSSM) with cross-attention and prototype-based similarity, and a second-pass ResidualSSM for error correction. The full pipeline achieves a 5-fold cross-validated macro-AUC of 0.890, compared to a Perch-only baseline of 0.739, representing a 15.1 percentage point improvement with only 59 labeled soundscapes available.