ISCA Archive - A compact low-cost speech recognition system (original) (raw)

A compact low-cost speech recognition system

A. Aktas, L. Glaßer, B. Kammerer, W. Küpper, M. Schlang

A compact low-cost speech recognition system for isolated words is presented. It yields a recognition rate of more than 95 % in a speaker dependent mode for telephone-quality speech. With little hardware expense a response time of less than half a second for a maximum vocabulary of 60 words is obtained. Real-time processing of the speech signal is achieved by restriction to the sign correlation for feature calculation and to binary energy information for word-boundary detection. A word is described by 12 feature vectors resulting from a linear temporal segmentation of the utterance. For classification, a pattern match with non-linear time alignment is performed. The system is realized on a standard eurocard pc board using an 8-bit microprocessor, 4 kByte ROM, and 8 kByte RAM. An on-board serial interface provides communication with a host computer or other external devices, while a communication field and an array of relays allow stand-alone operation of the recognizer.