A dynamic programming-based DUR recognizer implemented on analog devices' DSP

Abstract

This paper describes the use of Analog Devices' ADSP-2100 digital signal processor in discrete utterance recognition applications. The practicality of implementing conventional 100, 300, 500, 1000, 2000, and 5000-word speaker-dependent recognizers is explored. The processor's functionality is demonstrated via the coordination and execution of tasks which include d a t a acquisition, silence detection, feature extraction, amplitude normalization, and dynamic programming. A two-pass algorithm is presented which limits the computational burden of large vocabularies in the dynamicprogramming phase. On the first pass decimated prototypes and candidates are compared to obtain a set of best matches. A complete dynamic programming search is then performed on this set during the second pass. An adequate hardware implementation is shown to consist of a small general-purpose ADSP-2100 processor board and separate memory board.