Linguistic applications of speech technology require phone labels that accurately correspond to the phonetic signal. Mismatches between an ASR system's lexicon and speech acoustics can occur in situations of sociophonetic variation or phonological marginal contrast. We apply the recurrent neural network tool Phonet to Italian mid vowels, which are both sociophonetically variable and marginally contrastive. Phonet represents natural classes of sounds via interpretable vectors of phonological posteriors. Phonet was trained on 10 hours of Italian speech. Results show significant negative correlations between posterior probabilities for theoretically opposed classes like [open] and [close], alongside significant correlations of phonological posteriors to acoustics, e.g. values for [close] and vowels' first formant (F1). We demonstrate that Phonet’s phoneme recognition model generalizes to provide a better fit to vowel characteristics than phone labels derived from a G2P lexicon.