Novel Tonal Feature and Statistical User Modeling for Query-by-Humming

Abstract

This paper describes a query-by-humming (QbH) music information retrieval (MIR) system based on a novel tonal feature and statistical modeling. Most QbH-MIR systems use a pitch extraction method in order to obtain tonal features of an input humming. In these systems, pitch extraction errors inevitably occur and degrade the performance of the system. In the proposed system, a cross-correlation function between two logarithmic frequency spectra is calculated as a tonal feature instead of a difference of two successive pitch frequencies, and probabilistic models are prepared for all tone intervals existing in the database. The similarity scores between an input humming and musical pieces in a database are calculated using the probabilistic models. The advantages of this system are that it can obtain more appropriate tonal features than the pitch-based method, and it is also robust against inaccurate humming by the user thanks to its statistical approach. From experimental results, the top-1 retrieval accuracy given by the proposed method was 86.8%, which was more than 10 points higher than the conventional single pitch method. Moreover, several integration methods were applied to the proposed method with several conditions. The majority decision method showed the highest accuracy, and 5% reduction of retrieval error was obtained.