Title: Speech Recognition and Retrieval in Non-English Languages
Speaker: Jonathan Wintrode, US Department of Defense
Date: Thursday, April 9, 2009 3:30 - 4:30 pm
Location: DyDAn Center, CoRE Bldg, Room 431, Rutgers University, Busch Campus, Piscataway, NJ
Abstract:
We present an overview of speech recognition technology with an emphasis on building recognition systems for search and retrieval of audio in non-English languages. We begin with an overview of the theory underlying standard Hidden Markov Model (HMM) based speech recognition systems. We then look at the assumptions and limitations of the HMM-based framework in the context of both non-English languages and the retrieval task. Many of these limitations are more or less present in an English recognizer, but may become more pronounced in languages the speech community has not been working in for 40 years.
The second part of the talk looks specifically at retrieval of speech. We look at how the existing vector space model of information retrieval has been applied to the noisy output of speech recognizers. Topics covered with respect to both recognition and retrieval include: handling pronunciation and orthographic ambiguities, phonetic or sub-word recognition, and handling of out-of-vocabulary (OOV) terms.
See our new DyDAn Student-organized Seminar Series
Back to DyDAn Home Page
Back to DIMACS Home Page