Abstract
In this thesis, we propose new approaches to content-based music database retrieval. Traditionally, music is retrieved by keywords of song titles, performers’ names, or lyrics. In this study, we implemented a CBMR (content-based music retrieval) system that allows users to retrieve music based on some notes sung or hummed, or some rhythms tapped to the microphone.We proposed three different methods for content-based music retrieval based on acoustic inputs. The first one is “query by humming”, where the music notes of users’ acoustic input are segmented by amplitude level. The second one is “query by singing”, where note segmentation is achieved jointly by amplitude level and k nearest neighbor classification rule. The last one is called “query by tapping” in which the note duration is identified by amplitude change in the acoustic input. Once the test input and the reference songs in the database are transformed into the same mid-level representation, the similarity scores can be computed by variants of “dynamic time warping”. Then, the CBMR system returns the results according to the similarity scores. Currently there are 8544 reference songs in the database. The implementation details, performance evaluation, and the strength/weakness of each approach will be covered in this thesis.