The auditory neural representations of infants can easily be studied with electroencephalography using mismatch experimental designs. We recorded high‐density event‐related potentials while 3‐month‐old infants were listening to trials consisting of CV syllables produced with different vowels (/bX/ or /gX/). The consonant remained the same for the first three syllables, followed (or not) by a change in the fourth position. A consonant change evoked a significant difference around the second auditory peak (400–600 ms) relative to control trials. This mismatch response demonstrates that the infants robustly categorized the consonant despite coarticulation that blurs the phonetic cues, and at an age at which they do not produce these consonants themselves. This response was obtained even when infants had no visual articulatory information to help them to track the consonant repetition. In combination with previous studies establishing categorical perception and normalization across speakers, this result demonstrates that preverbal infants already have abstract phonetic representation integrating over acoustical features in the first months of life. Preverbal infants can compute automatically consonant representation, independently of the vocalic context. Infants share with adults a similar neural architecture suitable for computing phonetic representations from the first months of life.