Reference Annotations

Introduction

Shared data enable researchers to reliably compare their results with others. The Centre for Digital Music (C4DM) has a tradition of providing ground truth data for research, and our chord, onset and segmentation annotations have been used by many researchers in the MIR community. This year, we have focused on extending these resources in scope and quantity.

This web page accompanies a late-breaking paper at the 11th Conference on Music Information Retrieval [1].

The annotation data

The various annotations are available on the Isophonics Datasets Page but is advisable to continue reading below to understand the annotations and how they were derived.

The annotations published here fall into four categories: chords, keys, structural segmentations, and beats/bars. The aim is to allow more musically driven Music Information Retrieval by combining several features that are intrinsically linked. Of course, you can also use the annotations of one single musical feature separately.

We publish all data in two fundamentally different ways, namely a plain text format (with extension .lab or .txt, depending on the kind of data) and RDF. The RDF data is not as easily human-readable, but allows to use the data in a broader Semantic Web context. Note that while the text data is provided as separate files for every song and every feature, there is only one RDF file per song. Provided you have the original audio file, the RDF file can be directly opened in Sonic Visualiser, and all annotations for that song will be displayed.

We are proud to publish this data, but please be aware that it will never be exactly "correct", for several reasons:

  • music is always ambiguous, and different people will hear different things in the same music, so a label can only reflect the perceived truth of a person or a group of people,
  • the medium we present the annotations in is not perfect, and the syntax and more generally the method with which our features are annotated may still be improved, and finally
  • typos and other errors due to misinterpretations by the annotator (that go beyond the difference in perception as noted above) may occur.

The annotations come with "confidence labels" (see below, and in the individual RDF files), to tell you to which extent we think this last "typo" source of incorrectness is relevant.

Below you find more specific remarks to the individual annotation types for the text files. The RDF files should be self-explanatory to the expert user.

Chords

Comply with Chris Harte's syntax detailed in his 2010 PhD thesis [2] (which supercedes the 2005 ISMIR proposal paper [3]). The .lab files are whitespace delimited text files with three columns, corresponding to, onset time, offset time and chord label, respectively. Our confidence judgement:

  • The Beatles: checked several times by Christopher Harte and the MIR community, use with confidence.
  • Zweieck, Queen: checked by Matthias, use with moderate confidence.
  • Carole King: Have not been carefully checked, use with care.

Segmentations

These are structural segmentations, labelling verses, choruses, refrains etc.. The .lab files are whitespace delimited text files with three columns, corresponding to, onset time, offset time and segment label, respectively. Every segment starts at a bar boundary, and the section boundaries should be quite reliable. Parts (like a verse) that are partly repeated (maybe later in the song) but otherwise appear as one are usually treated as one, i.e. no subdivisions. The labelling however is not very consistent so far (note the "silence" labels), so using the labels to assess whether an algorithm correctly detects that two segments are "the same" may cause grief. Generally use with care.

Keys

These labels denote tonality regions in a song. The .lab files are whitespace delimited text files with three columns, corresponding to, onset time, offset time and key label, respectively. Generally use with care, especially the Beatles keys: the main key will always be annotated, but key changes may be omitted in some files.

  • Queen, Zweieck, Carole King: checked by Matthias Mauch, use with moderate confidence.
  • The Beatles: often only the main key is given. It may be the only key or it may not. Use with care.

Beats

The beat files come as whitespace delimited .txt files, which the first column denotes the beat time, the second column is the metric position. Bar labels can be derived from that by using only times at which the metric position is "1".

  • The Beatles: checked by Matthew Davies. Use with moderate confidence.
  • Zweieck: checked by Helena du Toit and Matthias Mauch. Use with moderate confidence.

Audio

We used the audio as provided on the following CDs:

  • Carole King: Tapestry, Ode Records, 4931802003
  • Queen: Greatest Hits I, Parlophone, 0777 7 8950424
  • Queen: Greatest Hits II, Parlophone, CDP 7979712
  • Queen: Greatest Hits III, Parlophone, 7243 52389421
  • The Beatles: Please Please Me, CDP 7 46435 2
  • The Beatles: With the Beatles, CDP 7 46436 2
  • The Beatles: A Hard Day’s Night, CDP 7 46437 2
  • The Beatles: Beatles For Sale, CDP 7 46438 2
  • The Beatles: Help!, CDP 7 46439 2
  • The Beatles: Rubber Soul, CDP 7 46440 2
  • The Beatles: Revolver, CDP 7 46441 2
  • The Beatles: Sgt. Pepper’s Lonely Hearts Club Band, CDP 7 46442 2
  • The Beatles: Magical Mystery Tour, CDP 7 48062 2
  • The Beatles: The Beatles (the white album), CDS 7 46443 8
  • The Beatles: Abbey Road, CDP 7 46446 2
  • The Beatles: Let It Be, CDP 7 46447 2
  • Zweieck: Zwielicht (contact for info or to purchase)

Acknowledgements

We'd like to thank our student annotators:

Eric Gyingy
Diako Rasoul
Felix Stiller
Helena du Toit
Vinh Ton
Chuks Chiejine

In the progress of labelling the chords, we used the following literature to verify our judgements:

Queen, Greatest Hits I, International Music Publications Ltd, London, ISBN 0-571-52828-7

Queen, Greatest Hits II, Queen Music Ltd./EMI Music Publishing (Barnes Music Engraving), ISBN 0-86175-465-4

Carole King, Tapestry, International Music Publications Ltd, London, ISBN 1-84328-452-9

Additionally, Alan W. Pollack's Notes on ... Series has been a great source of information.