Skip to content
All projects
CompetitionAI

BirdCLEF+ 2026

Identifying 234 Pantanal wildlife species from their calls in field recordings.

Role
Competitor
Year
2026
Illustration of a bird singing beside its spectrogram and audio waveform

What it is

A training and inference pipeline for the BirdCLEF+ 2026 Kaggle competition: multi-label classification of 5-second audio clips into probabilities for 234 species.

The problem

Field audio is noisy, overlapping and unevenly labelled, and the final inference has to run on CPU within the competition's time limit.

Architecture

Requests flow left to right. Hover a component for its role.

BirdCLEF+ 2026 architectureField recordings: 5-second clipsField recordingsLog-mel spectrograms: 128 melsLog-mel spectrogramsSpecAugment and MixupSpecAugment andMixupEfficientNet-B0 SED: attention poolingEfficientNet-B0 SED5-fold training: mixed precision5-fold trainingCPU inference: 234 probabilitiesCPU inference
Solid line: synchronous callDashed line: through a queueDashed box: external or infrastructure
  • Field recordings external, 5-second clips
  • Log-mel spectrograms service, 128 mels
  • SpecAugment and Mixup service
  • EfficientNet-B0 SED ai, attention pooling
  • 5-fold training worker, mixed precision
  • CPU inference service, 234 probabilities

Key decisions

  • Sound-event detection, not plain classification

    Attention pooling over time captures when a call happens within a clip and trains more stably than classifying the whole clip at once.

Highlights

  • Focal loss for rare species.
  • 18 automated tests over the pipeline.

Stack

PythonPyTorchtorchaudiotimmlibrosaaudiomentations