
Abstract
This study is a contribution to link the abstract phonological level to the acoustic signal level by identifying the main acoustic correlates for the distinctive feature set developed by Chomsky and Halle (1968). The acoustic features were extracted by the openSMILE toolkit from spontaneous speech data. For each distinctive feature a set of closely related acoustic features was derived by means of correlation-based feature selection. Based on the respective acoustic feature pools C4.5 trees and support vector machines for binary feature classification were trained. The classification performance ranged from 76 to 89% for vocalic features and from 78 to 93% for consonantal features. The methods proposed in this study can be of use to identify systematic speech signal correspondencies for phonological models and as a starting point for distinctive feature detection in speech recognition.
Item Type: | Conference or Workshop Item (Paper) |
---|---|
Form of publication: | Postprint |
Keywords: | distinctive features, acoustic features, feature selection, machine learning |
Faculties: | Languages and Literatures > Department 2 > Speech Science |
Subjects: | 400 Language > 410 Linguistics |
URN: | urn:nbn:de:bvb:19-epub-18045-5 |
Language: | English |
Item ID: | 18045 |
Date Deposited: | 22. Jan 2014, 11:46 |
Last Modified: | 04. Nov 2020, 12:59 |