At least two fundamental properties should be present in a network computing a phonetic representation: categorical perception and normalization across different utterances. Normalization processes were studied at birth by recording high density evoked potentials to strings of syllables in sleeping neonates. We compared the response to a change of phoneme when irrelevant speaker variation was present or absent. A mismatch response was recorded at the same latency in both cases, suggesting that relevant phonetic information was extracted from the irrelevant variation. Combined with our previous work showing that the mismatch response is sensitive to categorical perception in infants, this result suggests that a phonetic network like that of adults, is already present in the infant brain. Furthermore, efficient phonetic processing does not require attention.