This study investigates the perception of Cantonese lexical tones in sung syllables. Previous studies assumed that native listeners found the lexical tones unintelligible if the song was composed with low tone-melody correspondence, i.e., lyrics would sound odd and awkward in native ears. The degree of confusion and the cues that listeners use to understand tone information in sung syllabus have thus been overlooked. The present experiment aims to explore whether listeners can recognize the individual words produced in a song with low tone-melody correspondence. This involved 67 Cantonese speakers who identified six Cantonese tones from 42 syllables extracted from a song with lyrics sung in Cantonese with low tone-melody correspondence. Participants showed statistically significant correct perception of all six tones (overall accuracy rate = 29.85%; chance level = 16.67%), although T2, T5, and T6 were confused with other tones with accuracy rates near/below chance level. Participants also tended to misperceive other tones as T1. Furthermore, when T1 was sung at higher notes and T3, T4 and T6 sung at lower notes, they were more accurately identified. The results showed that even in a song with low tone-melody correspondence, some lexical tones can still be accurately identified, implying a complex interaction between tone and melody in perception.