ACTIVE SPEAKER DETECTION IN A SOCIAL HUMAN-ROBOT INTERACTION MAINTENANCE SYSTEM FOR SOCIAL ROBOTS

The main purpose of this paper is to apply a speaker detection algorithm to a human-robot interaction maintenance system. Social robotics places great emphasis on robots that are human companions, supporting medical facilities or other places of everyday use. Such robots are required to have the highest level of interaction in conversation with humans. This can be achieved by maintaining appropriate eye contact with the person you are talking to. Artificial intelligence solutions come to the rescue, in particular neural networks. The paper aims to present the implementation of the active speaker algorithm as a solution for maintaining human-robot interaction to a specially prepared system. This algorithm is mainly based on ResNET neural networks. The paper compares active speaker detection models in terms of their accuracy and execution time to find the most suitable model for the previously assumed system architecture. The obtained results show better accuracy of the TalkNet model with comparable model execution time. This makes the model ideal for a real-time human-robot interaction maintenance system. The proposed issues and solutions may attract more attention to improving social robot systems responsible for communication, which in the future may break the stereotype that people should spend as little time with robots as possible.