Acoustic Classification of Pig Behavior in Grow-Finish Operations: Toward Transport Monitoring

This research examined the possibility of transferring a fine-tuned model trained on local microphone data recorded in a private barn, as well as on open-source datasets, to new microphone devices, such as those built into a commercially available smartphone. A straightforward approach for comparing the model's predictions over one-minute intervals is used. Showing correlation of 0.4 to 0.53 between the four different microphones and the smartphone, while the microphones between each other show correlation between 0.68 and 0.87.
Not only the microphone characteristics, but also the interplay between positioning and complex background noise on farms lower those correlation coefficients. Nevertheless, reliable algorithms should work across different device types and not be prone to error from small location shifts.
In summary, the results show the ability of applying the model to smartphone audio data, achieving a moderate correlation between the predictions indicating a clear nonrandom meaningful relationship.
For further research, the different possibilities of audio settings in the smartphone and their effect on the model's performance should be investigated. Furthermore, the model performance should be evaluated on data collected from a pig transporter under real environmental background noise during transport.