Normalized Cuts for Singing Voice Separation and Melody Extraction
(M. Lagrange, L. G. Martins, J. Murdoch, G. Tzanetakis) - TSALP07

Data Sets and
Audio Results
H.W.P.S. Measure
Download and Run the Marsyas Peak Clustering Software

Download and Run the
Marsyas Peak Clustering Software

We are pleased to make available the software created as a part of this work. Here you will find information regarding downloading, compilation, and use of the "peakClustering" tool created using the Marsyas 0.2 framework.

Step 1: Download and Install the Marsyas 0.2 Framework.

  •  Marsyas is available on SourceForge at
  •  Installation instructions for Linux, MacOSX, and Windows
     are provided in the Marsyas 0.2 Users Manual provided below

Step 2: Use "peakClustering" (found in the /src directory) to analyse and synthesize peak clusters from independent or automatically-mixed pairs of audio files.

  •  The following explains the command-line usage of "peakClustering":
./peakClustering [OPTIONS] FILE
(Windows Visual Studio users, please consult the Marsyas Users Manual for project file information)
FILE Name of Audio File to be Processed

-i dir Input Directory Name
-o dir Output Directory Name
-q Duration Number of seconds of audio to analyze
-N file Audio file to be mixed with the target file before processing

-a Analyze from Audio File (.au, .wav, etc.)
-s Synthesize Audio From Selected Clusters
-c nbClusters Desired Number of Peak Clusters per Texture Window
-p pruning Number of Clusters to be Retained per Texture Window
-t type Peak Similarity Measure (Explained Below)
-P Save Clustered Peaks to File (.peak) (Mandatory for Further Off-Line Processing)

-A Compute Cluster Attributes from Peak File (.peak)
-S Synthesize Independent Audio for Each Cluster
Peak Similarity Measure ( -t [T_OPTIONS] )
Options are composed of any number of (attribute,normalization) pairs concatenated together

ATTRIBUTES a = amplitude
f = frequency
h = harmonicity
t = time
A = delta amplitude
F = delta frequency
NORMALIZATION o = unnormalized
n = normalize
l = use logarithmic scale
     (dB for 'a' or Bark for 'f')
b = normalize and use log scale
e.g. -t hoab

General documentation for the Marasyas 0.2 framework:
(please note that this documentation is a work-in-progress)

Marsyas 0.2 User Manual

Slides from the tutorial entitled, "MIR for Audio Signals Using Marsyas-0.2" presented at ISMIR 2006 (Victoria, B.C.):

ISMIR 2006 Tutorial Slides

Marsyas Logo