From a new acoustically transparent pneumotachograph mask, we simultaneously recorded aerodynamic (oral and nasal airflow) and acoustic data for 6 French male speakers, involving 3 oral and 3 nasal vowels out of logatoms (i.e. non words). A Convolutional Neural Network (CNN) trained on other acoustic corpora in French was tested on the data collected from the mask for the nasal/oral vowel distinction, with a 88% correct classification on average. We compared these CNN results with the nasal airflow extracted from all vowels of the corpora. Aerodynamic results showed a higher quantity of nasal airflow for the nasal vowels. However, for some speakers, distinction between nasal and oral vowels in terms of nasal airflow was less prominent, especially for /a/ vs /ɑ /, the 2 vowels for which the CNN have the least correct identifications. Finally, we discuss the discrepancies observed between aerodynamic data and CNN probabilities, and inter-speaker variations that can be approached by the CNN.