Deep Learning to Extract Laboratory Mouse Ultrasonic Vocalizations from Scalograms

We tested two techniques that can assist in the automatic extraction and analysis of laboratory mouse ultrasonic vocalizations. First, we substituted a Morlet-wavelet-based scalogram in place of the commonly used Fourier-based spectrogram. The frequencies in a scalogram are logarithmically scaled, which likely makes it better model a mouse's pitch perception based on the structure of its inner ear. Works that use the linear spectrogram are more likely to overemphasize frequency change at high frequencies, and so might misclassify calls when partitioning them according to frequency modulation. We showed that laboratory mice do indeed modulate their calls more when making high-frequency calls, which we would expect if they perceive pitch logarithmically. Secondly, we used "deep" convolutional neural networks to automatically identify calls within the scalogram. A convolutional neural network performs much like the animal visual cortex, and has enjoyed recent success in computer-vision and related problems. We compared the convolutional network to a conventional fully-connected neural network and the Song filter used by recent papers, and found that our convolutional network identifies calls with greater precision and recall.