Skip to Main content Skip to Navigation
Conference papers

Are IP initial vowels acoustically more distinct? Results from LDA and CNN classifications

Abstract : Past results have suggested that initial strengthening (IS) effects target the contrastive phonetic properties of segments, with a maximization of acoustic contrasts in initial position of strong prosodic domains. Here, we investigate whether IS effects translate into a better acoustic discriminability within the French oral vowels system. Discriminability is assessed on the basis of classification results of two types of classifiers: a linear discriminant analysis (LDA) based on the four formants frequencies, and a deep convolutional neural network (CNN) based on spectrograms. The test set includes 720 exemplars of /i, y, e, ɛ, a, x, u, o, ɔ/ (with /x/=/ø, oe/) produced in a labial context, either in intonational phrase initial (IPi) or word initial (Wi) position. Classifiers were trained using a set of 4500 vowels extracted from a large read speech corpus. Results show a better discriminability of vowels (overall better classification rate) in IPi than in Wi with the two methods. Less confusion in IPi is found between rounded and unrounded, and between back and front vowels, but not between the vowels along the four-way height contrast. Less confusion between peripheral and central vowels also expresses a maximization of contrasts within the acoustic space in IPi position.
Document type :
Conference papers
Complete list of metadatas

Cited literature [22 references]  Display  Hide  Download
Contributor : Cécile Fougeron <>
Submitted on : Tuesday, September 1, 2020 - 7:52:47 PM
Last modification on : Thursday, September 3, 2020 - 5:15:21 AM


Publisher files allowed on an open archive




Fanny Guitard-Ivent, Gabriele Chignoli, Cécile Fougeron, Laurianne Georgeton. Are IP initial vowels acoustically more distinct? Results from LDA and CNN classifications. Interspeech 2019, Sep 2019, Graz, Austria. ⟨10.21437/Interspeech.2019-2153⟩. ⟨hal-02927652⟩



Record views


Files downloads