Development of nonvoice dialogue interface for robot systems

In this presentation, a nonvoice dialogue interface scheme for robot systems to improve the flexibility of human‐robot interaction is proposed. A prototype system with a microphone in a plastic doll has been developed to examine its potential for distinguishing interaction noises and human utterances. It realizes a simple, one‐input one‐response interaction with perception of a user’s action. For example, the system can reply ‘‘Ouch!’’ when its body or head is beaten by a user. It is also able to make a response to a calling by the voice. This scheme would provide a new interaction style, unlike conventional spoken dialogue systems in which a user’s rough actions were treated as meaningless obstacles. It identifies the source of noises based on GMM (Gaussian mixture model) noise recognition [A. Lee et al., Proc. INTERSPEECH, 1, 173–176 (2004)]. Thirteen‐class GMMs were constructed from voices and noises like slapping or stroking the head, and knocking the body, which were recorded via trial testing the prototype system. It is possible to discriminate noises and utterances by comparing acoustic likelihoods from GMMs. The experiment investigates performances of acoustic feature vectors consisting of MFCC (mel frequency cepstral coefficients) to identify interaction noises.