LINGUIST List 19.2315
Mon Jul 21 2008
Diss: Phonetics: Tjalve: 'Accent Features and Idiodictionaries: On ...'
Editor for this issue: Hannah Morales
<hannahlinguistlist.org>
1. Michael
Tjalve,
Accent Features and Idiodictionaries: On improving accuracy for accented speakers in ASR
Message 1: Accent Features and Idiodictionaries: On improving accuracy for accented speakers in ASR
Date: 20-Jul-2008
From: Michael Tjalve <m.tjalveucl.ac.uk>
Subject: Accent Features and Idiodictionaries: On improving accuracy for accented speakers in ASR
E-mail this message to a friend
Institution: University College London
Program: PhD in Experimental Phonetics
Dissertation Status: Completed
Degree Date: 2007
Author: Michael Tjalve
Dissertation Title: Accent Features and Idiodictionaries: On improving accuracy for accented speakers in ASR
Dissertation URL: http://www.phon.ucl.ac.uk/research/phdabstracts/Tjalve07.pdf
Linguistic Field(s):
Phonetics
Dissertation Director:
Mark Huckvale
Dissertation Abstract:
One of the most widespread approaches to dealing with the problem of accentvariation in ASR has been to choose the most appropriate pronunciationdictionary for the speaker from a predefined set of dictionaries. Thisapproach is weak in two ways: firstly that accent types are more numerousand more variable than can be captured in a few dictionaries, even if theknowledge were available to create them; and secondly, accents vary in thecomposition and phonotactics of the phone inventory not just in whichphones are used in which word.
In this work, we identify not the speaker's accent, but accent featureswhich allow us to predict by rule their likely pronunciation of all wordsin the dictionary. Any given speaker is associated with a set of accentfeatures, but it is not a requirement that those features constitute aknown accent. We show that by building a pronunciation dictionary for anindividual, an idiodictionary, recognition accuracy can be improved over asystem using standard accent dictionaries.
The idiodictionary approach could be further enhanced by extending the setof phone models to improve the modelling of phone inventory and variationacross accents. However an extended phoneme set is difficult to build sinceit requires specially-labelled training material, where the labelling issensitive to the speaker's accent. An alternative is to borrow phone modelsof a suitable quality from other languages. In this work, we show that thisphonetic fusion of languages can improve the recognition accuracy of thespeech of an unknown accent.
This work has practical application in the construction of speechrecognition systems that adapt to speakers' accents. Since it demonstratesthe advantages of treating speakers as individuals rather than just asmembers of a group, the work also has potential implications for howaccents are studied in phonetic research generally.
|