For a QBH (query-by-humming) system, it is important to extract melody from various humming inputs (whistle, nasal, singing, and humming). We use a new method to classify the humming inputs into two types (whistle and non-whistle), and two different ways to extract fundamental frequency (FO). Two phase of cepstrum with zero-process is introduced to get proper fundamental frequency from the non-whistle type. With tracking the energy difference curve and pitch curve, assisted by cepstrum peak value curve, note segmentations are made. Rough experiments show that this approach is effective and efficient