A Study Of Melody Feature Extraction In Query By Humming System

In content based image retrieval methods, query by humming is a research hotspot. And Query-by-Humming(QBH) system based on Waveform files will be the main trend of the future. However, a waveform file is usually a mixture of strong accompaniment and singing voice. Accompaniment and singing voice have many overlaps in spectrum, which leads to a big interference to singing voice signal. So how to accurately extract the main melody is a very valuable research topic in the case of containing accompaniment for waveform files. This paper studies two kinds of Melody Extraction algorithm for waveform file and a Harmonic/Percussive Sound Separation (HPSS) algorithm. The main work is as follows:1. A melody pitch extraction algorithm based on Auditory Scene Analysis is studied in this paper.Firstly, we take an auditory peripheral processing for the input signal. The input signal passes through a128-filter gammatone filter bank, which can decompose the signal into many separated time-frequency units. Then we calculate the sum of autocorrelation sequence for each frame signal and establish a main melody pitch contour for the music. Secondly, we take an auditory segmentation and an auditory combination for the time-frequency units according to the characteristics of the auditory scene analysis. Finally, we determine the main melody pitch contour according to the constraint conditions of voice and singing voice.2. A melody pitch extraction algorithm based on Hidden Markov Models is studied in this paper.Firstly, we partitions and classifies an input into vocal and nonvocal portions. Then we take an auditory peripheral processing for vocal portions. Finally, we estimate the main melody pitch contour using HMM and statistical methods. Experiments show that the algorithm can accurately extracts the main melody pitch contour of the music.3. A Harmonic/Percussive Sound Separation (HPSS) algorithm is studied in this paper.A Harmonic/Percussive Sound Separation (HPSS) procession is increased before the two algorithms, one of which is based on auditory scene analysis, while the other is based on HMM. This procession can filter off part of the stringed instrument harmonic energy and increase the SNR (Signal to Noise Ratio) in mixed signals. Experiments show that the increasing of HPSS can effectively improve the accuracy of the main melody pitch extraction.