Transputer Implementation of the TRACE Model of Speech Recognition

Abstract : We describe the accomplishments of the five-year project to develop an extended version of the TRACE model of speech recognition and to port it to a Transputer platform. We begin with an over view of the project. We then describe the TRACE Model. TRACE is an artificial neural network which attempts to model the processing of speech from acoustic input to word extraction, in a manner consistent with what is known of human perceptual abilities. Following that, we discuss the work done to extend and develop the model in the direction of real- time performance with an extended phonemic inventory. This includes algorithm changes to TRACE, software for managing the Transputer processors, a new acoustic/phonetic front-end, debugging and testing tools, and graphics visualization software.