Speech Recognition Application for Pronunciation Training
Is Version Of
The Symposium provides undergraduates from all academic disciplines with an opportunity to share their research with the university community and other guests through conference-style oral presentations. The Symposium represents part of a larger effort of Iowa State University to enhance, support, and celebrate undergraduate research activity.
Though coordinated by the University Honors Program, all undergraduate students are eligible and encouraged to participate in the Symposium. Undergraduates conducting research but not yet ready to present their work are encouraged to attend the Symposium to learn about the presentation process and students not currently involved in research are encouraged to attend the Symposium to learn about the broad range of undergraduate research activities that are taking place at ISU.
The first Symposium was held in April 2007. The 39 students who presented research and their mentors collectively represented all of ISU's Colleges: Agriculture and Life Sciences, Business, Design, Engineering, Human Sciences, Liberal Arts and Sciences, Veterinary Medicine, and the Graduate College. The event has grown to regularly include more than 100 students presenting on topics that span the broad range of disciplines studied at ISU.
There is a large number of immigrants in the U.S. who speak English as a second language. Their spoken English may not be completely intelligible to both native and other non-native speakers of English due to errors in the production of specific sounds as well as in the prosodic features of pronunciation. This study used speech recognition (SR) technology to improve pronunciation abilities of non-native speakers of English. Ten participants were asked to read 20 sentences, 5 of which were not used in the experiment (these sentences served as control to ensure that SR training rather than just participation in the experiment accounted for any changes). Participants were instructed to read these sentences up to five times similarly to a native speaker model, which was provided to them. The extent of improvement, if any, in pronunciation was measured by tallying errors made by SR in recognizing speech produced by participants in successive production of the same sentences. There was an overall decrease in errors following training. A t-test for related samples showed a significant improvement from Time 1 to Time 5 (p=0.001).