Adverse environmental conditions cause a decrease of performance in automatic speech recognition. To cope with this problem, a real-time hardware module for robust distortion-insensitive word recognition has been developed. In addition, different coding algorithms for transmission and dialog announcements have been implemented. Thus complemented by appropriate speech processing software, this general-purpose unit provides all flexibility needed for the integration into systems like mobile telephone, voice-activated control of subsets and information systems.

Keywords: real-time implementation, robust feature extraction, background noise, neural net noise reduction, dynamic normalization, template adaptation