A Low Cost Real-Time Connected Word Speech Recognizer for Stock Trading and other Form-Entry Applications.

A speaker trained connected word recognizer has been designed for hands free form entry applications. A stock trading application comprising 78 vocabulary words and an 88 node grammar has been tested on this recognizer resulting in a word error rate of 1% in speaker trained mode. The recognizer runs on two IBM-PC compatible plug in cards. The main components are a 396f codec, an AT&T DSP32, 5 AT&T GSM's (graph search machines), and static RAMs. Surface mount technology could produce a single pc-card implementation. A training package has been designed for the pc which enables a single speaker to train the recognizer for his own voice. Software tools have been written for generation of a grammar for the given application. The hardware and software are easily extendable for larger vocabularies and grammars.