Cantonese, a syllabically paced, southern Chinese dialect, is a tonal language. A Cantonese syllable can have up to 9 different tone patterns which are lexically important. In this paper after reviewing major approaches to incorporating tone information into a large vocabulary continuous speech recognition (LVCSR) system, we propose two schemes to employ the tone information as a confidence measure for improving Cantonese LVCSR performance: either by guiding the stack decoder to search word lattice or by rescoring the N-best strings. It is shown the proposed schemes yield 4.2% and 8.0% relative improvement of Chinese character error rate, respectively.