High order structure in mouse courtship vocalizations

Many complex motor behaviors can be decomposed into sequences of simple individual
elements. Mouse ultrasonic vocalizations (USVs) are naturally divided into distinct syllables
and thus are useful for studying the neural control of complex sequences production.
However, little is known about the rules governing their temporal order. We recorded USVs
during male-female courtship (460,000 USVs grouped into 44,000 sequences) and classified
15 them using three popular algorithms. Modeling the sequences as Markov processes revealed
a significant temporal structure which was dependent on the specific classification algorithm.
To quantify how syllable misclassification obscures the true underlying sequence structure,
we used information theory. We developed the Syntax Information Score and ranked the
syllable classifications of the three algorithms. Finally, we derived a novel algorithm (Syntax
20 Information Maximization) that utilized sequence statistics to improve the classification of
individual USVs with respect to the underlying sequence structure.