Knowledge-Driven vs Data-Driven Methods for Filtering Acoustic Measures in Phonetics Corpora

Lancien, M., Adda-Decker, M. and Stuart-Smith, J. (2023) Knowledge-Driven vs Data-Driven Methods for Filtering Acoustic Measures in Phonetics Corpora. In: 20th International Congress of Phonetic Sciences (ICPhS), Prague, Czech Republic, 7-11 Aug 2023, pp. 3166-3170. ISBN 9788090811423

Abstract

This paper addresses the issue of vowel formant filtering for large scale phonetic analyses and evaluates an innovative data-driven method to remove spurious items. Automatic formant detection is error-prone due to formant jumps among other issues. A common solution is to adopt formant filters (FF) discarding tokens with measurements falling too far away from knowledgebased references. The proposed approach uses the Mahalanobis distance (MD) as a purely datadriven method. First, all vowel formant and duration values are used to compute vowel profiles. These learnt profiles supersede the need for reference values to carry out the filtering. We compare the two (knowledge-based vs data-driven) filtering approaches on the same dataset of French spontaneous speech. Results demonstrate the efficiency of MD filtering. The amount of filtered data is easily adjustable. Moreover, the data-driven status of the approach makes it well suited for less described languages.

• Item Type: Conference Proceedings
• Additional Information: The authors want to thank the Swiss National Science Fundation for the funds made available for this project, as well as the recorded speakers for the data we are using.
• Status: Published
• Refereed: Yes
• Glasgow Author(s) Enlighten ID: Stuart-Smith, Professor Jane and Lancien, Dr Melanie
• Authors: Lancien, M., Adda-Decker, M., and Stuart-Smith, J.
• Subjects: P Language and Literature > P Philology. Linguistics
• College/School: College of Arts & Humanities > School of Critical Studies ; College of Arts & Humanities > School of Critical Studies > English Language and Linguistics
• ISBN: 9788090811423
• Copyright Holders: Copyright © 2023 GUARANT International spol. s r. o.
• Publisher Policy: Reproduced under a Creative Commons License

Related URLs

• Organisation