In Voice Response Systems (VRSs) allowing the speaker-independent recognition of small vocabularies over the telephone network, a significant number of errors is due to faintness in speech/non-speech detection. In this paper, we firstly present a speech/non-speech detection evaluation method. This method helps to evaluate the speech/non-speech detection algorithm currently used by CNET. We then propose an improved adaptive version of this speech/nofl-speech detection algorithm. The evaluation results show great improvements.

Keywords: Speech/non-speech detection; Adaptive automaton Robustness; Voice Response Systems.