Discourse markers (DMs) are (chunks of) words stemming from the diachronic development of other parts-of-speech that tag the discourse's organization (ex. "well then", "innit"...). However, in synchrony, the formal accounts for the DM class vary from purely discourse-oriented definitions to models relying on a combination of lexico-grammatical and discursive information. We propose to bring new evidence into this debate by comparing the phonetic realizations of 4 DM types: stemming originally from adverbs, coordinators, subordinators and interjections. A discourse-only account would predict that the 4 types would be realized similarly, while a syntactic-discursive account predicts that subordinators would stand out, as they are less prone to syntactic independence. The analysis of various acoustic parameters (segment duration, F0, F1, F2 and HNR) in a finely-annotated 4-hour long corpus of French indicates that a hybrid approach may indeed be more accurate.