Variability of speech signals is known to reflect linguistic units' predictability, which is often measured with lexical frequency and contextual probability. Since the same aspects of the signals often serve other functions such as conveying lexical contrasts and marking phrasal boundaries, it is interesting to examine to what extent the link between predictability and phonetic variability is constrained and manifested differently across different dimensions of phonological contrasts. This question motivates the present study’s focus on F0 realizations in Taiwan Southern Min, a language with a rich inventory of lexical tones. From a corpus of spontaneous speech, the relationship between predictability measurements and the realization of different tonal categories is analyzed with Generalized Additive Models. Results show that predictability effects generally do not neutralize critical F0 contrasts (e.g., F0 peak between high-falling and low-falling tones, F0 mean between high-level and low-level tones). For some tonal pairs, there is a trend of lower predictability (i.e., higher surprisal and informativity) correlating with a larger F0 difference. These findings shed light on how the relationship between phonetic variability and predictability is modulated by the maintenance and enhancement of lexical contrasts in speech production.