Deep convolutional network for animal sound classification and source attribution using dual audio recordings.

Area of Science:

• Bioacoustics
• Animal Behavior
• Machine Learning

Background:

• Analyzing animal vocalizations is crucial for understanding behavior.
• Current methods for classifying animal calls can be limited by environmental noise and data quality.
• Marmoset monkey vocalizations offer insights into social structures and communication.

Purpose of the Study:

• To develop an end-to-end convolutional neural network (CNN) for classifying marmoset call types and sources.
• To enable reliable classification in noisy environments using multi-stream audio data.
• To enhance the capacity for analyzing marmoset vocalizations in research settings.

Main Methods:

• An end-to-end feedforward CNN was designed and trained.
• The network utilizes two streams of audio data.
• Raw spectrogram images were used as input for classification.
• Training data included recordings from captive marmosets with imperfect labels.

Main Results:

• The CNN reliably classified both the type and source of marmoset calls.
• Accurate classification was achieved even in noisy environments.
• The network processed data in a single pass using raw spectrograms.

Conclusions:

• The developed CNN significantly increases data analysis capacity for researchers studying marmoset vocalizations.
• This approach allows for non-invasive data collection in the home cage and group-housed animals.
• The study demonstrates the potential of AI in bioacoustics research.