An adaptive speech recognition system operating in a remote time-shared computer environment

This paper describes the general approach, hardware, and computer program structure used in a speech recognition system coupled to a time-shared computer facility. The tasks that the system performs are split between the hardware, which is used for the complex analog processing tasks, and the software, which can more readily perform the pattern recognition and data file management chores. It is linked to a time-shared computer system via a voice-grade telephone line. An RCA 752 video terminal or a Model 35 teletype is used for entering control messages and to receive the output from the recognition process. The resulting system can be adaptively trained to handle a small but arbitrary vocabulary by any individual. The system operates in real time, with a speed of response less than a few seconds. The level of accuracy is comparable to that which has been obtained for similar medium-scale systems. The system is characterized by: 1) A novel configuration-with division of hardware and software. 2) Nonprivileged and real-time operation in a time-shared environment. 3) Low data rate-110 baud and voice-grade telephone lines can be used. 4) Use of context to segment large vocabularies and achieve lower error rates and reduce computation. These innovations point to the first steps toward a practical and economical speech recognition system.