Speaker trained isolated word recognition on a large vocabulary of words

Publisher Website
Google Scholar
Abstract
Most isolated word recognition systems have been tested on small to moderate size vocabularies, e.g., 10 to 200 words. For continuous speech recognition, however, vocabulary sizes of greater than 1000 words are required for a wide variety of tasks. An important baseline performance measure for such systems is the recognition accuracy of the system on isolated words drawn at random from the vocabulary. As such, we have measured the performance of an LPC-based recognizer on an 1109-word vocabulary (Basic English with a set of scientific terms appended) operating in a speaker trained mode. Each of six talkers (three male, three female) trained the system using the robust training procedure of Rabiner and Wilpon. Following training, each talker spoke the entire 1109-word vocabulary four times over the course of a month of testing. Recognition accuracies as a function of talker, vocabulary size, and word position were measured. Results indicate that accuracy is a strong function of both vocabulary size, and choice of vocabulary words from the Basic English vocabulary.

This publication has 0 references indexed in Scilit: