An automatic speech detection architecture for social robot oral interaction

摘要

Social robotics have become a trend in contemporary robotics research, since they can be successfully used in a wide range of applications. One of the most fundamental communication skills a robot must have is the oral interaction with a human, in order to provide feedback or accept commands. And, although text-to-speech is an almost solved problem, this isn't the case for speech detection, since it includes a large number of different conditions, many of which are literally unpredictable. There are quite a few well established ASR (Automatic Speech Recognition) tools, however without providing efficient results, especially in less popular languages. The current paper investigates different speech detection strategies via the utilization of the Sphinx-4 open-source library. The first is a way to incorporate languages for which no acoustic or language model exists (Greek in our case), following the grapheme-to-phoneme concept. The speech detection model is evaluated using audio captured from a NAO v4 robot, a difficult task due to the high levels of included noise, thus denoising techniques are investigated as well.