Normalized Cuts for Predominant Melodic Source Separation

Abstract

The predominant melodic source, frequently the singing voice, is an important component of musical signals. In this paper we describe a method for extracting the predominant source and corresponding melody from "real-world" polyphonic music. The proposed method is inspired by ideas from Computational Auditory Scene Analysis. We formulate predominant melodic source tracking and formation as a graph partitioning problem and solve it using the normalized cut which is a global criterion for segmenting graphs that has been used in Computer Vision. Sinusoidal modeling is used as the underlying representation. A novel harmonicity cue which we term Harmonically Wrapped Peak Similarity is introduced. Experimental results supporting the use of this cue are presented. In addition we show results for automatic melody extraction using the proposed approach.

An iterative approach to monaural musical mixture de-soloing

“…Given this new matrix and the other parameters in Θ, we could possibly compute the separated signalsv (1) andm (1) . However, since we modified HF 0 , the estimated parameters are no longer optimal, especially for WK , and a second estimation taking into account this new parameter matrix is necessary and improves the separation as shown by the results in section 4.…”

“…Our database is composed of 3 subsets: (A) the SiSEC 2008 development set for the "professionally produced music recordings" separation task 1 , (B) some songs from Ozerov and Lagrange's private database ([4] and [1]) and (C) publicly available songs by S. Hurley, under Creative Commons licence. C is further divided into a pitch contour annotated set C1 and its complementary set C2.…”

“…It is also gaining interest for a number of Music Information Retrieval (MIR) applications such as music and drum transcription since it can, as a pre-processing step, ease the indexing of complex polyphonic signals [1], [2]. The aim of BASS is to extract separated contributions (or sources) from a mixture by exploiting their differences in terms of spatial location and/or time-frequency (or timbral) content.…”

“…They are then separated using a Wiener time-frequency mask. On the other hand, approaches like [1], [5] or [6] rely on sinusoidal models and unsupervised techniques to label several groups of sinusoids as belonging to either of the expected sources. The use of a sinusoidal model however significantly impairs the subjective quality of the results.…”

Towards effective singing voice extraction from stereophonic recordings

“…Mono approaches utilize such diverse methods as pitch detection and amplitude modulation [3], source-adapted models [4], and normalized cuts [5]. On the other hand, most stereo methods make use of the Interchannel Intensity Difference (IID) in addition to the Interchannel Time Difference (ITD) or the Interchannel Phase Difference (IPD) [1], [6][7][8].…”