BovineTalk: machine learning for vocalization analysis of dairy cattle under the negative affective state of isolation.
Authors: Gavojdian Dinu, Mincu Madalina, Lazebnik Teddy, Oren Ariel, Nicolae Ioana, Zamansky Anna
Journal: Frontiers in veterinary science
Summary
# Editorial Summary: Machine Learning Analysis of Cattle Vocalisations as Welfare Indicators Identifying reliable, non-invasive markers of negative emotional states in dairy cattle remains a significant gap in on-farm welfare assessment, particularly as precision livestock farming technologies become increasingly integrated into production systems. Researchers developed and validated two computational approaches—deep learning and explainable machine learning models—to classify high-frequency and low-frequency cattle vocalisations and identify individual animals by voice, using the largest curated dataset to date of lactating dairy cows experiencing visual isolation stress. The deep learning framework achieved 87.2% accuracy for low-frequency calls and 89.4% for high-frequency calls, whilst individual cow identification reached 68.9% accuracy; the explainable machine learning model performed comparably at 89.4% and 72.5% respectively, with the latter approach offering greater interpretability for practical application. Given that high-frequency vocalisations are strongly associated with negative affective states and dairy cows routinely encounter significant stressors throughout their production cycle, these validated acoustic classification systems could be incorporated into automated on-farm monitoring protocols to flag welfare concerns in real time. The explainability of the machine learning model is particularly valuable for farriers, veterinarians, and farm managers seeking to understand *why* the system has identified distress, rather than operating as a 'black box' decision tool.
Read the full abstract on PubMed
Practical Takeaways
- •Vocalization analysis via automated acoustic monitoring could become a practical on-farm welfare indicator for dairy cattle, potentially flagging stress or negative emotional states before behavioural signs are obvious
- •The ability to identify individual cows by voice opens possibilities for precision health monitoring without visual identification, useful in larger herds or poor lighting conditions
- •High-frequency calls should be monitored as a potential early warning signal for isolation stress or other negative challenges in dairy production systems
Key Findings
- •Machine learning models achieved 87.2–89.4% accuracy in classifying high- and low-frequency cattle vocalizations using deep learning and explainable ML frameworks
- •Individual cow voice recognition was possible with 68.9–72.5% accuracy, demonstrating vocalization contains individuality information
- •High-frequency open-mouth calls are associated with negative affective states and long-distance communication in dairy cattle
- •Visual isolation-induced vocalizations provide a non-invasive, animal-based indicator suitable for integration into precision livestock farming tools