Voicing variations in French obstruents: Distribution and acoustic quantification

This study investigates phonetic variation of voicing correlates in French obstruents. A controlled set of words was extracted from a large broadcast news corpus providing 378 word-initial singleton obstruents /t, d, k, s/. An expert investigation, by eye and ear, of the productions revealed that phonologically voiced obstruents are phonetically voiced most of the time (96%), while a large number of phonologically voiceless obstruents are produced with a partial or complete phonetically voiced constriction, more frequently so for stops (74% /t/, 61% /k/) than for fricatives (30% /s/). In order to acoustically quantify our observations and establish reliable metrics for future larger-scale studies, 13 acoustic metrics were applied and tested. Out of these, three metrics were found to be particularly effective in accurately classifying our obstruents into the manually defined categories: an energy difference measure relative to the following vowel, an unvoiced/voiced frame ratio, and consonant duration.