Using Mahalanobis Distance to Filter Erroneous Vowel Features in Less-Resourced Languages: Application to Quebec French

Lancien, M., Stuart-Smith, J. and Adda-Decker, M. (2023) Using Mahalanobis Distance to Filter Erroneous Vowel Features in Less-Resourced Languages: Application to Quebec French. In: 20th International Congress of Phonetic Sciences (ICPhS), Prague, Czech Republic, 7-11 Aug 2023, pp. 3181-3185. ISBN 9788090811423

Abstract

A major challenge in building shareable datasets for phonetic studies consists of maximising data collections while minimising the errors involved, in particular when automatic processes come into play (e.g. labeling, formant detection...). Automatically extracting formants from large amounts of speech frequently produces artifacts, due to formant jumps, alignment problems, or noise. One solution is to use formant range filters. However this requires prior knowledge about the vowels, such as formant range information. Here we propose to use the Mahalanobis distance to remove erroneous values relying only on the labeled speech data. Our study is conducted on a Quebec French corpus including more than 170 k tokens of 16 vowel types. Results show that the proposed method can complement the threshold-based filter approach. Furthermore, it can be used autonomously for undocumented languages to eliminate erroneous values. The approach also makes it easy to adjust the degree of filtering.

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

Related URLs

• Organisation