Combined spectral and speech features for pig speech recognition.

Area of Science:

• * Acoustics
• * Computer Vision
• * Animal Science

Background:

• * Pig vocalizations are crucial indicators of health, hunger, and emotional states.
• * Traditional speech recognition methods using only spectral features may limit accuracy in pig sound classification.
• * A need exists for advanced methods to accurately interpret pig sounds for timely health interventions.

Purpose of the Study:

• * To develop and validate a more accurate pig sound classification method.
• * To enhance the monitoring of pig health and welfare through acoustic analysis.
• * To establish a comprehensive dataset for future research in pig vocalization analysis.

Main Methods:

• * A novel pig sound classification method leveraging both signal spectrum (spectrograms) and time-domain audio features.
• * Implementation of a parallel network structure to process and integrate dual feature inputs.
• * Selection and combination of the best-performing network model and classifier.

Main Results:

• * Achieved a classification accuracy of 93.39% on the pig sound classification task.
• * Reached an Area Under the Curve (AUC) of 0.99163, demonstrating high model performance.
• * Established a dataset of 4,000 pig sound samples across four categories.

Conclusions:

• * The proposed dual-feature method significantly outperforms traditional single-feature approaches for pig sound classification.
• * This research provides a robust foundation for automated pig health monitoring systems.
• * The established dataset will facilitate further advancements in animal acoustics and welfare research.