Connected-word recognition using the NeXT workstation

Abstract

In the field of man/machine interface, connected-word recognition systems are becoming essential for efficient interaction. Various algorithms have been implemented which combine accuracy with computational efficiency and required storage. The authors present a real-time implementation of a connected-word speech recognition using the NeXT workstation as the DSP platform. The NeXT workstation offers the advantages of having integrated DSP, real-time, and interface tools readily available for the development environment. The on-board DSP32 offers the signal process capacity for real-time speaker recognition. The Motorola 68030 is used to handle the system-wide data flow. A 3rd-party A/D unit (Digital Ears) is used as the input device. The connected-word recognition algorithm combines a max/min coding parameter with a varying-order Markov model. The resulting system is speaker-dependent with accuracy exceeding 95%. Vocabulary sizes of up to 100 words have been developed. >