QBH systems are designed to identify the most similar songs in database using hummed query. The process begins by capturing a hummed query on the user side, continues with its transcription using pitch detection algorithms and ends with user being presented the results of search algorithms. The goal of this thesis was to examine existing QBH systems in order to find the most optimal one for use in web application EtnoFletno, which was followed by its implementation. Algorithms YIN and probabilistic YIN were considered and tested for query transcription phase of the target system. Several search algorithms were implemented and tested as well, including DTW, Edit Distance, Spring, SMGT and SMBGT. Transcription and search algorithms had to be optimized for usage in EtnoFletno, hence the testing database contained Slovenian folk songs. Final results show that the transcription algorithm probabilistic YIN was better than its predecessor YIN and algorithm SMBGT outperformed all other search algorithms. It is also shown in the results that algorithm SMBGTs parameters should be used in two different predefined ways considering users singing skills.
|