VEMUS Music Benchmark Data Set

Overview

VEMUS Music Benchmark Data Set was compiled during the three years (2005-2008) of VEMUS project. VEMUS stands for "Virtual European Music School" and was a project funded by the European Commission under the Information Society Technologies (IST) Programme of the Sixth Framework Programme (FP6). The VEMUS resulted in an open, highly interactive, and networked multilingual music tuition framework for popular instruments and a set of innovative pedagogically-motivated e-learning components addressing individual, classroom and distance learning settings. Among all partners of the Consortium, Kungliga Tekniska högskolan (Sweden) and UAB "Balteck" (Lithuania) were two major contributors to this Data Set.

VEMUS Music Benchmark Data Set was used for the development and testing of audio recognizers embedded into VEMUS system. This page is intended to open VEMUS Data Set to the music recognition research community as a Benchmark Data Set for comparing different music recognition approaches and algorithms.

License terms

By downloading/using this Data Set or parts of it your agree to comply with the following terms and conditions:

1. The Data Set may be used for research purposes only.

2. The Data Set, its parts and copies thereof, may not be sold, leased, published or distributed to any third party without written permission from the VEMUS consortium.

3. All publicly released research results obtained using this Data Set must acknowledge VEMUS Data Set was used.

4. Neither the VEMUS consortium nor any individual Partner of the consortium shall be held liable for any errors in the contents of the databases and damage arising from the use of the databases.

5. The VEMUS consortium may unilaterally modify the conditions of use for this Data Set.

Structure

VEMUS Music Benchmark Data Set consists of:

Music recordings
Annotations and MIDI files
Software

Music recordings contain solo performances of various woodwind/brass instruments performed by beginners and professional students. These are actual field recordings and may be occasionally superimposed with a speech and/or environmental noise.

Every audio file has an associated annotation file which is prepared by the human listener. An annotation file contains time data of every note onset as well the sequence of note pitches (C#4, D5 notation). Annotation files are text files stored in TextGrid (short version) format. They can be easily visualized with Praat software developed by Paul Boersma and David Weenink at the Institute of Phonetic Sciences, University of Amsterdam. Every audio file has an associated MIDI file which essentially represents the same information as contained in a TextGrid file but which can be played as well.

Software is designed to perform benchmark testing, i.e. to compare selected annotation files of the Data Set (reference data) to the annotations produced by some music recognizer. Comparison is based on the dynamic programming approach for aligning two note sequences. Comparison results for VEMUS off-line audio recognizer are provided for illustration purposes. Binary executables as well as the source code of the software can be downloaded and used under the terms of GNU public license.

Music recordings, annotations and MIDI files

Instrument	Recordings	MB	Audio	MIDI	TextGrid
Clarinet	90	459
Euphonium	27	209
Flute	100	560
Recorder	143	247
Saxophone	42	158
Trombone	34	205
Trumpet	12	44
Tuba	4	29

Software for benchmark tests

Windows (compiled by bcc32 v5.5.1 Borland free compiler)	compare_win32.exe
Linux	N/A
MacOS	N/A
Source code (ANSI C++)	compare_source.zip
Sample data (VEMUS Offline music recognizer)	Recognized scores
Sample results	Summary Details
Documentation

Related publications