DECODING POULTRY WELFARE: AN INTEGRATED MACHINE LEARNING AND NLP FRAMEWORK FOR VOCALIZATION ANALYSIS
Date
2025-08-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Poultry vocalizations serve as valuable, non-invasive indicators of health, stress, and overall welfare in farm environments. This thesis presents an integrated machine learning framework combining classical acoustic analysis and natural language processing (NLP) to decode these vocal cues. The first approach uses features like MFCCs, spectral contrast, and zero-crossing rates with models such as Random Forest, LSTM, and TabNet for classifying health, behavior, and stress. The second pipeline transcribes raw audio using Wav2Vec2.0 and applies BERT-based sentiment and linguistic analysis to detect welfare-related vocal shifts. Ensemble models showed strong generalization across welfare contexts, while NLP models uncovered subtle stress-induced vocal patterns. Scalable preprocessing with ThreadPoolExecutor enabled efficient handling of large datasets. These methods support the development of an on-farm, real-time monitoring system for early stress detection, reduced manual labor, and improved animal welfare—contributing to more ethical, sustainable, and productive poultry farming practices.
Description
This thesis investigates how vocal behavior in poultry can be used as a window into their health and emotional state, offering a novel, non-invasive approach to welfare monitoring. Rooted at the intersection of bioacoustics, machine learning, and natural language processing (NLP), the research builds a dual-framework system: one focused on acoustic signal processing and statistical classification, the other on semantic and emotional interpretation of vocal patterns.
What makes this work distinct is its integrative use of AI — Using traditional classifiers like LSTMs and Random Forests and advanced transformer-based models such as BERT and Wav2Vec2.0. Beyond improving classification accuracy, the thesis emphasizes interpretability, scalability, and real-world deployment potential, addressing key gaps in current poultry welfare research.
By enabling continuous, automated monitoring of bird vocalizations, this work supports more ethical and efficient farming, with broader implications for smart agriculture, animal behavior research, and sustainable food systems.
Keywords
Machine Learning (ML), Artificial Intelligence, Poultry Vocalizations, Semantic Analysis, Natural Language Processing (NLP), Animal Welfare and Sustainability, Acoustic Sensing