SqueakOut: Autoencoder-based segmentation of mouse ultrasonic vocalizations.

Area of Science:

• Bioacoustics
• Computational Neuroscience
• Animal Behavior

Background:

• Mouse ultrasonic vocalizations (USVs) are crucial for social communication.
• Accurate segmentation of USVs from spectrograms, removing background noise, remains a significant challenge in bioacoustic analysis.

Approach:

• Developed SqueakOut, a lightweight, fully convolutional autoencoder utilizing a MobileNetV2 backbone with skip connections and transposed convolutions.
• Trained SqueakOut on a novel dataset of 12,954 annotated spectrograms specifically for mouse USV segmentation.
• Employed stochastic data augmentation and a hybrid loss function for robust segmentation across diverse recording conditions.

Key Points:

• SqueakOut achieves a high Dice score of 90.22 for supervised USV segmentation, significantly outperforming existing methods like VocalMat (63.82 Dice score).
• The model's lightweight architecture (4.6M parameters) makes it efficient for practical applications.
• The publicly released dataset and SqueakOut implementation will foster further research in mouse vocalization analysis.

Conclusions:

• SqueakOut offers a highly accurate solution for segmenting mouse USVs from spectrograms.
• This advancement facilitates more precise analysis of mouse communication patterns and enables novel classification methods.
• The availability of the dataset and tool promotes reproducible research and accelerates discoveries in the field.