A real-time speech recognition architecture for a multi-channel interactive voice response system

Abstract

Achieving reliable implementation of a real-time speaker independent speech recognizer through the telephone network is a challenging research problem. Besides requiring the selection of a suitable algorithm, which takes into account both real-time and accuracy constrains, it requires also adequate hardware architecture and optimized software. This paper presents a dedicated multiprocessor DSP architecture for telecom applications. Based on ADSP-21060 SHARC DSPs, it is the kernel of a digital interactive voice response (IVR) system that is connected to a digital switch through the primary CCITT standard time division multiplexing line of 2.048 Mbps. We attempt to show how a multi-DSP based hardware can be designed for a specific problem in telecommunication, along with the implementation of automatic speech recognition (ASR) to the digital IVR system.