Methods and Data
Project scope
This repository provides an end-to-end audio classification pipeline for detecting animal vocalizations in wildlife recordings. It covers preprocessing, feature extraction, augmentation, training, evaluation, and prediction. Although developed for chimpanzee vocalizations, the workflow can be adapted to other species when labeled recordings are available.
Dataset summary
Training data was collected at the Mefou Primate Sanctuary in Cameroon. Cross-environment evaluation was performed at Chimfunshi Wildlife Orphanage Trust in Zambia.
Two training configurations are supported:
- Sanctuary only: real chimpanzee vocalizations and background recordings from Cameroon.
- Sanctuary plus synthetic: the same data augmented with synthetic examples that mix Cameroonian vocalizations with Zambian background noise.
Training sample counts
| Dataset | Chimpanzee samples | Background samples |
|---|---|---|
| Sanctuary | 17,921 | 74,163 |
| Synthetic | 68,757 | 97,149 |
Step summaries
Preprocessing and condensation
The preprocessing workflow starts from Raven annotations and can optionally apply a condensation step to shorten long recordings into denser candidate-vocalization streams for annotation. A detailed description of the condensation algorithm, its parameters, and example usage is available on the dedicated condensation page.
Feature extraction
The repository provides both handcrafted acoustic features for SVM models and multi-channel spectrogram features for CNN models. The dedicated feature extraction page contains the original detailed description of the feature families, preprocessing, parameterization, and command-line usage.
Classification
The classification stage supports SVM and CNN workflows. The dedicated classifier page summarizes data segmentation, feature selection, CNN architecture, augmentation, training configuration, and the standard training, evaluation, and prediction commands.