Machine learning approaches for tracking changes in marmoset vocalizations and social behaviors

Abstract

Common marmosets are highly social and vocal primates. They display social vocal accommodation (SVA), a form of vocal learning involving modification of call types to match their acoustic properties to those of their social partners. This ability, along with speech-like vocalization properties, makes marmosets interesting animals to study the mechanisms underlying vocal learning and the functions thereof. Marmoset SVA has been previously studied using a handful of pre-selected acoustic features commonly employed for vocal analyses in other species. Such features may not capture all the acoustic changes taking place, suggesting the need for a more comprehensive approach. Using time-series analyses and hierarchical machine learning classifiers, I first showed that source identities could be determined from marmoset vocalizations with high accuracy. Next, by employing multiverse analyses to track changes in marmoset vocalizations and social behavior during SVA, I showed that such an approach provides additional information about the phenomena by comparing the various quantification methods. My analyses confirmed that the extent of SVA is the highest in close contact calls, suggesting trade-offs between SVA and preserving individual identities. However, most of the behavioral measures did not agree with each other nor correlate with the extent of SVA, suggesting the need for improved behavioral data acquisition techniques. Finally, I showed the possibility of using machine learning to synthesize marmoset vocalizations to test the role of SVA in signaling social closeness. These approaches will assist in providing detailed insights into the vocal and behavioral changes that occur during vocal learning in animals.