Query By Singing And Humming System

With the rapid development of high capacity storage and digital music technology, how to search music with its content information becomes a more and more important problem. Query by singing and humming system is one way to search music with its content. Different from those traditional music searching engines, which require too much effort to classify and label different kinds of music, the query by singing and humming system makes it possible for users to find the required music in a fast and convenient way.This thesis proposes a query by singing and humming system. Users query by singing or humming, and the proposed system uses the query to search in a MIDI music database and return the required song back to the user in the form of a top-N song list. Our work focuses on two parts of the system: 1) we proposed a new feature extracted from note sequence: melody histogram, represent the note sequence with this fixed dimension feature, and introduced the bit-vector-indexing (BVI) to the QBSH system, to index the music database, using melody histogram, so we can filter song segment candidates efficiently. 2) We studied the recursive alignment (RA) algorithm and made 3 improvements on it: change the distance measure, using mismatch penalty between pitch contour and note sequence, and offsetting the note sequence to match notes and pitch contour. These improvements increase the query accuracy efficiently.We use MIREX2008 QBSH corpus to evaluate our query by singing and humming system and the system achieved a mean MRR of 0.937. Also, comparison experiments are performed, revealing that BVI indexing and improved RA algorithm's contribution on filtering song segment candidates and system accuracy performance.