Audio signal analysis of cow vocalizations in Python’s Librosa with gradient boosting algorithms to identify abnormal sound signatures in lame cows

Objectives

Develop methods to record high-quality audio data from individual cows, investigate platforms to preprocess the audio signals, and build machine learning tools to identify abnormal vocalization.

Materials and methods

Raw audio data was collected from individual cows in dairy farms using the Zoom H4n Pro Handy Recorder. This series of Zoom recorders has sophisticated capabilities for field recording, such as XY microphones, preamplification modules, and enhanced noise cancellation. Each audio recording was captured for around 30 seconds and saved in a .wav file format within a local server. All the saved data were initially screened to discard the audio recordings of poor signal quality. Thereafter, data labeling was performed in the presence of an expert to label the audio recordings into different categories. The next stage was data preprocessing which included the steps of label encoding, feature scaling and database split (into Test, Train and Validation Sets). All models were built using python3. For our model development, we used Librosa which is a powerful Python package for music and audio signal analysis. [...]