Understanding speech requires the construction of phonetic representations while abstracting from specific sound features. To understand different speakers of varying pitches of voice, loudness, or timbre, categorical phoneme information needs to be rapidly extracted from dynamic, changing speech input. The present study demonstrated a genuine MMN to tokens of /a/ and /i/ vowels varying in pitch of voice and amplitude envelope when they occurred infrequently among the respective other vowels. These data indicate that the speech perception system pre-attentively extracted the F1/F2 formant information despite the language-irrelevant variation in the sound input.