Development of a voice control interface for navigating robots and evaluation in outdoor environments

In this paper the development of a prototypic mobile voice control for navigating autonomous robots within a multi robot system is described. As basis for the voice control a hidden markov model based speech recognizer with a very low vocabulary of 30 words is utilized. It is investigated how many training samples for a markov model are required for a normal operation of speaker-dependent speech recognition. Therefore, hidden markov models were developed successively in parallel with an own training data corpus containing finally 2290 utterances from 12 speakers. Within the successive development of acoustical models and training corpus, the work revealed details about how many speakers are necessary to achieve an acceptable degree of speaker independence. We focused on an evaluation of the speech recognizer in adverse outdoor environments. The evaluation ranges from almost calm conditions of about 39 dB up to very adverse noise conditions of 120 dB. It is investigated whether a small vocabulary attenuates the noise vulnerability and in how far an increase of speaking volume can compensate noises of different intensity. The voice control was tested in outdoor environments and aspects of its usage are described.