This paper presents a new finding and discusses some of its theoretical implications. The finding is that phonetic categorization, including major class membership, is entirely predictable from phonotactic biases in 3 Brazilian Portuguese word databases. The predictor parameters are log frequencies of 'VC, 'CV and V'CV sequences consisting of the 7 stressed vowels combined with the 19 onset consonants, plus the 5 pre-stressed vowels. Correct vowel categorization arises through discriminant function analysis of 'VC and 'CV data. Correct consonant categorization arises through discriminant function analysis of V'CV data. Results are consistent across databases and, thus, suggest that statistical biases in the lexicon can be stable enough to code phonetic categories. These findings have a bearing on the issue of the relationship of phonotactics to phonetics.