In this thesis, a song recognition system is proposed. Song recognition is one of many types in pattern recognition where not much research has been performed.
The system is composed of five stages : pitch extraction, tone/tempo estimation, dynamic time warping, fuzzy feature matching, and fuzzy integral. Dynamic time warping is used for correct matching of the estimated tone and tempo. Fuzzy matching and fuzzy integral are applied to the system for handling any ambiguities.
In the proposed system, song recognition is performed without regard to the key and tempo of the song. Evaluation of the songs is performed based on the results from fuzzy feature matching. The score of the song can also be extracted from the results of the system.
An experiment is performed to evaluate the proposed song recognition system. Ten songs are selected for song recognition from a sample data set of hundred songs. Each song is sung three times by five different male voices. The recognition rate using the first estimate is 88.7% and 91.3% with the second estimate included.