Voice Prints: Individual Signatures and Recognition in Mice

It is widely taken for granted that we humans can recognize many different individuals just by the sound of their voice. Individual recognition is an extraordinarily difficult challenge. It took many years to develop computer programs capable of individual voice recognition (also called speech recognition or speaker identification), and because of improvements in artificial intelligence, called "machine learning", such tools are now available and increasingly used for biometric security. Most species use vocalizations for communication, and some animals also use vocalizations for individual recognition (`individual vocal recognition`), such as for parent-offspring recognition and mate recognition. However, few studies have identified the acoustic features used for individual vocal recognition and no studies have solved how such recognition works. The first aim of our project is to investigate individual vocal signatures and individual recognition in wild house mice (Mus musculus). House mice emit complex ultrasonic vocalizations (or USVs) during courtship and other social interactions, but their communication functions are not understood. We found that the USVs of wild mice contain individual distinctive characteristics, which are consistent over several weeks, and we will now test whether mice use these individual signatures in USVs to recognize individuals. The second aim of our project is to use machine learning to develop a program that can recognize individual mice by their vocalizations. This is an interdisciplinary collaboration between behavioral biologists and acoustic engineers, and we recently engineered special computational tools for detecting USVs and for classifying different types of USVs. We now aim to develop a new and improved tool for classifying USVs, and a new tool for identifying individual mice by their USVs. We will train a machine (convolutional neural network) using many recordings of individual mice and then assess its performance under different conditions. Thus, our research will help to understand the functions of the ultrasonic vocalizations of house mice, and determine whether their calls mediate individual vocal recognition. If successful, our computational tools for studying mouse USVs will greatly facilitate research on mouse USVs by making it possible to assign USVs to specific individuals in social interactions.