A Simple and Lightweight Algorithm for Social Robot Speech Turn Taking

Abstract

Simple, but effective, social robot voice-based interaction designs are possible with only rudimentary speech analysis. Robust full analysis, including syllable, word, and meaning extraction, is still an open research problem, with existing solutions being computationally expensive (and thus a power drain) while suffering from high error rates. Instead, we note that it is sufficient for many social interactions for a robot to simply distinguish when a person starts and finishes talking, and indeed argue that designing social robot interactions around such a constraint may – in the short term – result in more robust behaviors. We introduce a simple solution, using standard lightweight signal processing techniques (i.e., involving the derivative of the audio’s RMS value), that detects the beginning and end of speech utterances, including a preliminary evaluation. We envision that this simple, easy to implement algorithm may be useful for researchers aiming to simply and quickly implement basic robotic speech turn taking on low-capability or power-constrained robot devices. Further, the approach can support innovation in simple conversation with social robots.