MCM algorithm cuts false positives in bird species classification from overlapping audio

Bird species classification from field recordings is notoriously difficult due to overlapping vocalizations and incomplete species labels. A team led by Yuzhu Wang tackles this by first applying source separation as a preprocessing step to disentangle overlapping bird songs. They employ an ensemble of two separation models—the recurrent FTRNN and the transformer-based TF-Locoformer—both trained with mixture invariant training (MixIT). However, separation errors can introduce false positive detections for species that are not actually present. To counter this, the authors propose mixture-constrained max pooling (MCM), which takes the predicted probability for each species from each separated channel and the original mixture, then clips the per-channel value to not exceed the mixture's probability. This simple yet effective clipping prevents overly confident false positives from dominating the final aggregation.

Experiments on two real-world bird song datasets show that the separator ensemble significantly improves recall (true positive gain) for overlapping species compared to classification without separation. Yet it also introduces false positive gain for absent species—a problem MCM directly addresses. Across multiple evaluation metrics, MCM outperforms standard max pooling (which simply takes the maximum across channels) by reducing false positives while retaining true positive improvements. The work demonstrates that careful post-processing of separated audio can unlock better multi-species detection without requiring perfect separation. Accepted at IWAENC 2026, this research is particularly valuable for ecological monitoring where automated recognition of multiple species from a single recording is critical.