Utilizing synthetic training data for the supervised classification of rat ultrasonic vocalizations

Area of Science:

• Bioacoustics
• Animal Communication
• Machine Learning in Biology

Background:

• Rodent ultrasonic vocalizations (USVs) are crucial for social behavior and communication.
• Manual analysis of USVs is labor-intensive and time-consuming.
• Existing machine learning methods for USV analysis have limitations in accuracy and training data requirements.

Purpose of the Study:

• To compare the performance of two convolutional neural networks (CNNs), DeepSqueak (DS) and VocalMat (VM), in detecting and classifying rat USVs.
• To evaluate the impact of using synthetic USVs to augment training data for the VM CNN.
• To assess the potential of these automated methods for laboratory use.

Main Methods:

• Comparison of human expert performance against DS and VM CNNs for USV detection and classification.
• Evaluation of VM CNN performance with and without synthetic USV data augmentation.
• Analysis of rat USV audio recordings.

Main Results:

• VocalMat (VM) CNN outperformed DeepSqueak (DS) CNN in both USV identification and classification accuracy.
• Augmenting VM training data with synthetic USVs significantly improved its classification accuracy.
• The enhanced VM CNN achieved performance comparable to human experts.

Conclusions:

• Automated analysis of rodent USVs using CNNs is feasible and efficient.
• VocalMat, particularly with synthetic data augmentation, offers a promising tool for USV analysis in research settings.
• This approach can reduce the manual workload and improve the consistency of USV analysis.