A music retrieval system based on the extraction of non trivial recurrent themes and neural classification
B. Colaiocco, Francesco Piazza · 2004
In this paper we propose a new approach for music features extraction used for fast content-based retrieval of songs from a suitable built database. The algorithm discovers, by an interaction with the program manager, the refrain of type O MIDI songs and builds a database in which musical features and text information such as title, author, genre, etc. are stored. A neural network architecture is trained only by the features from the refrain of all the songs in the database belonging to an appropriate sub-class, and performs the retrieval in query-by-humming problem kind. Elman recurrent neural nets are used. Experimental results show the effectiveness of the proposed approach.