Logo TU Vienna  Logo IFS Information & Software Engineering Group
Vienna University of Technology
Institute of Software Technology and Interactive Systems
Information & Software Engineering Group

Music Information Retrieval

Logo Music Information 
Retrieval at TU Vienna IFS
  [Topics] [Projects] [Downloads] [People] [Publications] [Press] [Events]  

Architecture

Metadata

Library

Comparison Algorithms

Interface

Downloads

Screenshots

 

A Comparative Query by Example System

Lukas Maczejka, Christian Seidl (lukas.[lastname]@compucare.at, c.[lastname]@vision05.at)

This project was concerned with the development of a framework for analysing and evaluating MIDI feature extraction and comparison algorithms. A number of algorithms has already been implemented and included within the framework.

 

System Architecture

The system consists of three interconnected main modules. Various metadata is read from MIDI sources and stored in a meta library. Depending on the format of the distinct meta data multiple algorithms for comparison are implemented, representing the numeric difference of two extracted features. Both the library subsystem and the comparison subsystem are designed to allow for easy extension.

 
 

Metadata

Metadata is stored with an unique identifier within the database (as a list of meta files within a folder). The system provides interfaces for storage and retrieval of these data sets. Another interface is provided for the selection and application of a comparison algorithm to a set of similar meta data. MIDI preprocessing is provided to allow for simple feature extraction and to minimise possible errors.

Metadata is stored per input MIDI in a file containing a header and the following information for each extracted feature:

  • 1: Feature name
  • 2: Number of lines
  • 3-N: Extracted data

    Implemented feature extraction algorithms:

  • Parson's Code (Up/Down Sequence)
  • Note Duration (Duration of single notes)
  • Note Histogram (Note histogram by note identifier)
  • Lempel Ziv - Longest Sequence (Adaption of the LZ78 algorithm for extraction of the longest sequences)
  • Lempel Ziv - Most common sequences
  • Note Sequence
  • MIDI Statistics

  •  
     

    Library

    The library subsystem manages a list of meta data and provides support for comparison of these meta files. Additional MIDI sources can easily be added to the library. After extraction all of the features are stored in a file within the managed library folder.

     
     

    Comparison algorithms

    A comparison algorithms compares a set of two meta data instances, resulting in the similarity (or difference) of these two sets. The similarity is numerically expressed between 0 and 1, where 1 describes identical features and 0 denotes a completely different feature set.

    Implemented comparison algorithms:

  • Dice’s Coefficient (Bi-Gram comparison of two strings)
  • Euclidean Distance (Comparison of numerical vectors)
  • Hamming Distance (Comparison of two strings with identical length)
  • Levenshtein Distance (Comparison of two strings with distinct length)

  •  
     

    Interface

    A user interface was developed for ease of access to the system functionality. The interface provides library management tools as well as tools for the comparison of data stored in the library. It is designed to support the user (or developer) in the task of testing and evaluating the applicability and quality of different audio feature extraction algorithms.

     
     

    Downloads

    Documentation: Ein Komparatives Query by Example System (in German)

    Source code: Java-Source

     

    Screenshots

    img1 img2 img3

     

     

     

     
     
    top
    last edited 22.10.2009 by Lukas Maczejka, Christian Seidl