Melody extraction from polyphonic music based on the amplitude relation

Abstract

Most of the melody extraction methods only focus on the temporal continuity of the frequency, and seldom consider the temporal correlation of the frequency amplitude, which would result in the poor effect of melody extraction for many algorithms. Consequently, a new main melody extraction method based on the correlation of the frequency amplitude is proposed in the paper. Firstly, an equal loudness filter is used to enhance the ear-sensitive frequency band in music. The STFT is used to convert the spectrum and the phase vocoder is used to correct the frequency and amplitude. Then, lots of most salience frequency points near spectral peaks are selected as pitch candidates by a new salience function based on the correlation of the amplitudes of adjacent frames, and the perceived pitch is reverse-reasoned by a pair of high frequency points. Some pseudo fundamental frequency points are filtered out by detecting the number and distribution of their harmonics. Next, the pitch with the greatest salience, the pitch with the greatest amplitude and the pitch with the most harmonics are selected for creating contours in each frame. After the creation of contour, we analyze the distribution of the amplitude in each contour and clip the fragments with smaller amplitude for determining the start point and the end point of the melody. Finally, the contours with the smaller mean salience and amplitude are removed and the main melody is identified when there is more than one contour simultaneously. The experimental results show that the proposed method can effectively extract the main melody from polyphonic music.