Files in this item



application/pdfECE499-Fa2015-chang.pdf (221kB)
(no description provided)PDF


Title:Gamesourcing Mismatched Transcription
Author(s):Chang, Yu Yao
mismatched transcription
transcribed speech
transcribed speech from non-native speakers
Abstract:Transcribed speech is an essential resource to develop speech technologies for different languages of the world. However, native speakers of most languages of the world may not be readily available online to acquire transcribed speech. The goal of this research is to explore the possibility of acquiring transcriptions for speech data from non-native speakers of a language, referred to as mismatched transcriptions. The two main problems tackled in this work are: 1) How do we motivate non-native speakers to provide transcriptions? 2) How do we refine the mismatched transcriptions? Firstly, we design a novel game that facilitates the collection of mismatched transcriptions from non-native speakers. In this game, players are prompted to listen to sound clips in a foreign language and asked to transcribe the sounds they hear to the best of their abilities using English text. The misperceptions by the non-native speakers are modeled as a finite memory process and implemented using finite state machines. The mismatched transcriptions are further refined using a series of finite-state operations. The main contributions of this thesis are as follows: 1) Creation of a streamlined game for crowdsourcing transcriptions for speech data from non-native speakers. 2) Algorithms that process the resulting mismatched transcriptions and provide the closest sounding English words. 3) Experiments describing various modifications to the above-mentioned algorithms and results showing their effect on the accuracy of the English words that are produced as output.
Issue Date:2015-12
Date Available in IDEALS:2016-02-18

This item appears in the following Collection(s)

Item Statistics