A Study Of Query By Humming System With Large Music Database

Music retrieval is a new hot topic on Content-based retrieval. As a branch of music retrieval, Query-by-Humming (QBH) provides a brand new way of retrieval for users. However, the humming audio may contain noise, and there may be a lot of difference in pitch and rhythm compared with the audio of the target song. Thus, it’s a worthy research subject to design a QBH system with large music database.The main work of this paper:1. Propose a new QBH system based on Deep Learning.There are two key problems of QBH:The first one is that the melody feature is not stable. This paper uses Deep Belief Network (DBN) to process melody feature to solve this problem. The second one is that the melody feature needs fuzzy matching. This paper achieves neighbor retrieval by clustering the melody feature.To be specific, in the period of building database of songs, the system extracts the melody feature of the songs and does clustering. Then, the system trains DBN model using clustering results as labels. Then, the system does feature transformation by DBN model and builds indexes. In the period of online retrieval, the system extracts melody feature of the humming audio and finds the according cluster by indexes and considers the melody feature in this cluster as candidates set. At last, the system sorts the candidates and returns the results.Experiments proof that the QBH system based on DBN makes the feature more stable and improves the accuracy compared to the QBH system based on Locality-sensitive Hashing.2. Design a QBH system with large music databaseFor large music database, this paper uses distributed index and paralleled retrieval to solve the problems of huge index and parallel query.To be specific, this paper designs and analyzes a QBH system with large music database. In order to store huge index, the system uses distributed index. In order to process paralleled query, the system uses the strategy of paralleled retrieval.Experiments proof that the QBH system with big music database can improve the speed of retrieval while the retrieval precision is stable. Distributed index makes it easy to expand database.