In the present study, the distinctive tones of Mandarin were found to coarticulate with adjacent tones in running speech. However, the amount of deviation of a tone from its canonical form due to coarticulation varied depending on the nature of the tonal context. In a context where adjacent tonal values agree (a "compatible" context), the deviation was relatively small. In a context where adjacent tonal values disagree (a "conflicting" context), the deviation was much greater, sometimes even to the extent of changing the direction of a dynamic tone. To examine the perception of coarticulated tones, naturally produced words and phrases were used as stimuli for tone identification. However, their semantic information was removed through waveform editing while all the tonal information contained in the signal remained intact. Identification of tones in the compatible context was highly accurate with or without the original tonal context. Tonal identification for the conflicting context remained accurate only when the tones were presented with the original tonal context. Without the original context, i.e., in isolation, correct tone identification dropped below chance. When the original tonal context was altered, listeners compensated for the altered context as if it had been there originally. It thus seems that, in tone perception, listeners compensate for variations due to coarticulation. Nevertheless, even with the presence of the original context, perceptual identification was better for the compatible context than for the conflicting context, indicating that variation due to coarticulation is not always completely compensated for by listeners.