Abstract
Karaoke singing is a popular activity in many countries. Karaoke machine provide dynamic lyrics to help user singing, but user may encounter a problem in querying a song. So far when user queries a song, user must recall the song names then lookup it in a song list given by the Karaoke machine. Compared to query by lookup system, query by singing/humming system is a more convenient way to make a query. Query by singing/humming system get a melody singing/humming by user, and then compare this melody to the melodies in the database to recognize the songs which have the melody similar to the melody given by user. Finally, the system output the recognized songs to user as reference to make a query. If we can add the query by singing/humming system into Karaoke machine to work with query by lookup system, we can improve the fluency. The database adapted by this research has all sentence onsets tagged, we will take advantage of this feature to recognize repeating melody fragment within a song and remove the unnecessary melody fragment. By this way, we can improve the recognition speed. By studying the cases which the system failed to recognize them correctly, we found that the tonality may vary during the user’s singing/humming period. So we fragment the singing/humming clip into sections and then do transposition with different sections, respectively. We also do some experiments to verify the benefit from this method.