This paper describes a retrieval system that enables a user to hum part of a melody as a query to obtain the name of a desired song from an audio database. The combined use of two musical features, tone transition and tone distribution, enhances the accuracy of the retrieval system. These features are represented as high dimensional vectors to facilitate the retrieval process. Similarity retrieval using indices of these feature vectors increases the speed of the system. This paper presents the techniques utilized in the system. Experimental results are also reported based on hummed tunes of casual singers.
展开▼