Files in this item



application/pdfECE499-Sp2019-hoffer-sohn.pdf (1MB)Restricted to U of Illinois
(no description provided)PDF


Title:Automatic speech recognition for low-resource languages and dialects
Author(s):Hoffer-Sohn, Yda
Contributor(s):Hasegawa-Johnson, Mark
Subject(s):Automatic Speech Recognition
Dialect Variation and Adaptation
Abstract:Most widely spoken languages have numerous dialects or accents which can vary in degree of mutual intelligibility between speakers and automatic speech recognition systems alike. Oftentimes it is difficult or costly to build a specified dataset for training an automatic speech recognizer, and in some cases, it is not feasible to do so. In order to increase the usability of speech recognition software trained on available data, this thesis research explored methods of training ASR models on dialectal speech. Common methods for training ASR models on low-resource and dialectal speech often rely on the idea of “there’s no data like more data”, and employ various methods for data augmentation, multi-lingual and cross-lingual training, or gleaning data from less obvious sources. In contrast to these common approaches, this thesis discusses a proposed phonetic mapping layer added to a high-resource single language speech recognition system to branch between a standard dialect and a regional dialect with a set of common alternate phoneme pronunciations.
Issue Date:2019-05
Date Available in IDEALS:2019-06-14

This item appears in the following Collection(s)

Item Statistics