Research On Melody Extraction In Polyphonic Music

Melody extraction in polyphonic music signal is a technology which enables users to interact with music directly. It has a close connection with audio signal processing and music information retrieval (MIR). The technology, which is usually called "Melody Extraction","Audio Melody Extraction","Predominant Melody Estimation" or "Predominant Frequency Estimation",aims at extracting the frequency sequence from melody automatically.This paper is based on multi-frequency music extraction method and consider the fact that the music signal is the combination of vocal and instruments.In order to obtain the frequency of the vocal, it's necessary to extract the candidates of vocal frequency in frequency domain and then make use of short-time features to discriminate the predominant frequency, and use long-time features to distinguish the vocal frame so that a melody frequency sequence will come out finally. This paper has analyzed and studied the following three aspects:1. Put forward three kinds of discriminative features between the predominant frequency of vocal and instrument in harmonic domain.This paper analyzes the differences between human voice and instrument based on the pronunciation theory and concludes their potential distinctions. Then,this paper conducts a large number of statistical analysis experiments on the harmonic structure domain of voice, instrument and polyphonic music and puts forward three kinds of discriminative features: harmonic structure amplitude,harmonic structure frequency error and harmonic peak numbers.2. Put forward three melody discrimination methods based on harmonic structure.This paper puts forward several new melody discrimination methods by making use of the combination of harmonic structure features and logistic regression algorithm. Then, this paper compares and analyzes the advantages and disadvantages of these methods.3. Study and accomplish the function the vocal frame determination in preprocessing stage.This paper makes use of the statistical features in MIR field and logistic regression classifier and applies it to the preprocessing stage of melody extraction,resulting in the decrease of false rate and the increase of the accuracy.