SincNet-Based Speaker Identification for Robotic Environments with Varying Human–Robot Interaction Distance

Abstract

As human–robot interaction (HRI) becomes increasingly significant, various studies have focused on speaker recognition. However, few studies have explored this topic in the specific environment of home service robots. Notably, most existing research relies on databases composed of English-language data, while studies utilizing Korean speech data are exceedingly scarce. This gap underscores the need for research on speaker recognition in robotic environments, specifically using Korean data. In response, this paper conducts experiments using a speaker recognition database tailored to the Korean language and set in a robotic context. The database includes noise generated by robot movement as well as common environmental noise, accounting for variable distances between humans and robots, which are partitioned accordingly. The deep learning model employed is SincNet, with experiments conducted under two settings for the SincNet filter parameters: one with learnable parameters and the other with fixed values. After training the model with data collected at varying distances, performance was tested across these distances. Experimental results indicate that SincNet with learnable parameters achieved a peak accuracy of 99%.