Introduction
In modern high-intensity poultry farming, ensuring animal welfare is a significant challenge. Traditional monitoring relies on periodic manual inspections by experts, which are subjective, labor-intensive, and provide only intermittent snapshots of flock health. Acoustic monitoring presents a non-invasive, continuous alternative. However, broiler vocalizations are complex and change significantly as birds age—their peak frequency can drop from 3.6 kHz to 1.5 kHz in just five weeks. This research introduces an automated acoustic detector capable of classifying these vocalizations throughout the broiler life cycle.
Methodology
The researchers developed a convolutional neural network (CNN) architecture optimized for acoustic classification. The model processes audio inputs as log-mel spectrograms, which capture the spectral characteristics of the birds' calls.
Key methodological innovations include:
- Multi-Task Learning: The model simultaneously predicts vocalization type and the bird's normalized age. This helps the network account for the natural "pitch shift" that occurs during growth.
- Meta-Model Architecture: A 5-branch meta-model was used during fine-tuning, allowing the integration of features at different levels of complexity, similar to Inception modules.
- Advanced Denoising: A nonstationary noise suppression algorithm was applied to filter out farm background noise, such as ventilation systems and machinery.
- Dataset: The system was trained on a library of 2,559 labeled recordings covering four vocalization types: distress calls, pleasure notes, short peeps, and warbles.
Results and Discussion
The system achieved a balanced accuracy of 87.9% across five classes (four vocal types plus background noise). Notably, the model demonstrated high sensitivity to welfare-critical sounds, reaching a 97.2% recall rate for distress calls.
The inclusion of age estimation proved vital; the model can estimate a bird's age with a mean absolute error of approximately 2 days. Furthermore, ablation studies showed that denoising was the most critical preprocessing step, improving performance by 7.1 percentage points by preventing the model from learning "false clues" from background machinery.
Conclusion
This CNN-based detector offers a scalable and objective tool for real-time welfare monitoring. While challenges remain regarding overlapping sounds and generalization across different farm environments, the high accuracy in identifying distress signals suggests significant potential for automated health alerts in the poultry industry. The model is computationally efficient enough for real-time edge deployment on standard hardware.
Summary updated