The increasing needs of content-based automatic indexing for large musical repositories have led to extensive investigation in musical sound pattern recognition. Numerous acoustical sound features have been developed to describe the characteristics of a sound piece. Many of these features have been successfully applied to monophonic sound timbre recognition. However, most of those features failed to describe enough characteristics of polyphonic sounds for the purpose of classification, where sound patterns from different sources are overlapping with each other. Thus, sound separation technique is needed to process polyphonic sounds into monophonic sounds before feature extraction. In this paper, we proposed a novel sound source separation and estimation system to isolate sound sources by maximum likelihood fundamental frequency estimation and pattern matching of a harmonic sequence in our feature database.