Thank you for your interest in the McGill Billboard annotations! We are proud to announce a 2.0 release, with many improvements over the original, including new annotations, a number of corrections to errors in the existing annotations, and more accurate estimates of timing.
In order to facilitate the best possible use of these data in the future, we have made them available legally under a CC0 license, but we ask that users follow scholarly norms in any public-facing work based on upon these data by citing the following ISMIR paper:
or Ashley Burgoyne’s dissertation:
Please note that although the SALAMI project  contributed the structural metadata to the McGill Billboard annotations, we gathered (and funded) the chord annotations independently from the SALAMI project, hence the distinct citation.
This release contains the annotations and audio features corresponding to the first 1000 entries from the random sample of Billboard chart slots as presented at ISMIR 2011, plus the additional 300 entries that were used to evaluate audio chord estimation for MIREX 2012. The set includes annotations and features for 890 slots, as we were unable to acquire audio for every entry in the sample, and comprises 740 distinct songs, as due to the nature of the sampling algorithm, some slots correspond to the same song. Training algorithms that assume independent, identically distributed data (as most do) should retain the duplicates. We will release annotations for the remaining 700 entries progressively over the next couple of years in order to ensure that there are unseen data available for evaluating algorithms at MIREX or related events.
Since the original release, we have been able to complete a number of annotations that had originally encountered problems. Some of these new annotations are a closer match to the target sample than what was available at the time we released the original McGill Billboard annotations, and as such, a small number of entries refer to a different song in version 2.0 than they did in 1.x releases. We recommend that all users replace any previous editions of these annotations with version 2.0.
This release is also split into multiple files, to reflect the needs of different users. Users may choose their preferred archive format (XZ, BZ2, or GZ) and extract all their relevant archives into a single directory. The result will be a
McGill-Billboard directory with a subdirectory for each annotated entry in the sample, each subdirectory containing the relevant annotations and features.
Most users will want to download the index to the dataset:
The index is a CSV files with the following columns:
Sample entries for which we were unable to obtain audio or an annotation also appear in the index, but with the id, chart_date, and target_rank columns exclusively.
The complete annotations – chords, structure, instrumentation, and timing – are available from these links:
The annotations files are named
salami_chords.txt to reflect the fact that they contain both SALAMI-style structural annotations and the McGill Billboard chord annotations.
Each annotation begins with a header including the title of the song (prefixed by
# title:), the name of the artist (prefixed by
# artist:), the metre (prefixed by
# metre:), and the tonic pitch class of the opening key (prefixed by
# tonic:). Similar
tonic comments may also appear in the main body of the annotations, corresponding to changes of key or metre. In some cases, there is no obviously prevailing key, in which case the tonic pitch class is denoted
The main body of each annotation consists of a single line for each musical phrase or other sonic element at a comparable level of musical structure. Each line begins with a floating-point number denoting the timestamp of the beginning of the phrase (in seconds) followed by a tab character. There are special lines for
silence at the beginning and end of the audio file and a special line for the
end of the piece. The other lines continue with a comma-separated list of elements among the following.
More detail on the structural annotations is available from Smith et al. . The McGill Billboard annotations replace the lower level of structural annotations from this reference (lowercase letters) with chord annotations.
The chord annotations are simplified to the beat level. All chord symbols follow the standard that Harte et al. presented at ISMIR 2005 and used in MIREX ever since , with a few additions to the shorthand to facilitate the richness of these annotations:
1 for unharmonised bass notes,
5 for power chords, and
min13 for the corresponding chords in traditional jazz notation. An additional pseudo-chord type of
1 denotes bass notes with no chord on top. To save space, repeated chords are denoted with a dot instead of the full chord name. To further save space, bars containing a single chord on all beats list the chord symbol only once; likewise, in quadruple metres (4/4 or 12/8), bars with only two chords and the change on the third beat list those two chords with no dots. For brief changes of metre, the metre may appear in parentheses at the beginning of the bar rather than as a full metre comment.
Two non-chord symbols may appear within bars. For passages that were too musically elaborate to merit beat-level chord annotations, annotators sometimes filled the bar with an asterisk (*). For brief pauses of arbitrary length (often a single beat), annotators added a bar with the special annotation
Bas de Haas has written Haskell tools for parsing and manipulating the McGill Billboard annotations , which are available from Hackage:
Users who are only interested in automatic chord recognition may prefer to download HTK-style LAB files for the chord annotations instead, which contain only onset times, offset times, and the chord labels, as used for the audio chord estimation task in MIREX:
Note that only the first chord of each phrase was time-aligned by a human. The timings for all other chords are linearly interpolated assuming a constant tempo for each phrase. This constant-tempo assumption is remarkably robust for the McGill Billboard sample: less than one percent of all possible eighth-note positions (tatums) in the sample are more then 10 percent faster or slower than the average tempo of the songs to which they belong .
For convenience, we also have LAB files with chord labels simplified to the vocabularies that will be used for evaluating chord estimation in MIREX 2013:
Although we cannot distribute the original audio due to copyright, we have two feature sets available. Users interested in chord recognition may want the non-negative-least-squares chroma vectors and tuning estimates from the Chordino VAMP plugin :
These archives contain
tuning.csv for each annotated single. We used the default settings for the plugin with the exception for a rolloff of 1 percent, the plugin authors’ recommendation for pop music.
Researchers of many kinds may find the Echo Nest features helpful. We have recomputed these with the Echo Nest Analyzer version 3.1.4:
If you are interested in audio features other than these, please contact us. So long as the features are non-invertible and the computational load is sane, we are happy to provide custom features upon request.
Please e-mail any questions, comments, or bugs to Ashley Burgoyne at firstname.lastname@example.org.
John Ashley Burgoyne, Jonathan Wild, and Ichiro Fujinaga, ‘An Expert Ground Truth Set for Audio Chord Recognition and Music Analysis’, in Proceedings of the 12th International Society for Music Information Retrieval Conference, ed. Anssi Klapuri and Colby Leider (Miami, FL, 2011), pp. 633–38, http://ismir2011.ismir.net/papers/OS8-1.pdf.
John Ashley Burgoyne, ‘Stochastic Processes and Database-Driven Musicology’ (PhD diss., McGill University, Montréal, Québec, 2012), http://digitool.Library.McGill.CA:80/R/-?func=dbin-jump-full&object_id=107704&silo_library=GEN01.
Jordan B. L. Smith, J. Ashley Burgoyne, Ichiro Fujinaga, David De Roure, and J. Stephen Downie, ‘Design and Creation of a Large-Scale Database of Structural Annotations’, in Proceedings of the 12th International Society for Music Information Retrieval Conference, ed. Anssi Klapuri and Colby Leider (Miami, FL, 2011), pp. 55–60, http://ismir2011.ismir.net/papers/PS4-14.pdf.
Christopher A. Harte, Mark B. Sandler, Samer A. Abdallah, and Emilia Gómez, ‘Symbolic Representation of Musical Chords: A Proposed Syntax for Text Annotations’, in Proceedings of the 6th International Conference on Music Information Retrieval, ed. Joshua D. Reiss and Geraint A. Wiggins (London, England, 2005), pp. 66–71, http://ismir2005.ismir.net/proceedings/1080.pdf.
W. Bas de Haas and John Ashley Burgoyne, ‘Parsing the Billboard chord transcriptions’ (Technical report UU-CS-2012-18, Utrecht University, the Netherlands, 2012), http://www.cs.uu.nl/research/techreps/repo/CS-2012/2012-018.pdf.
Matthias Mauch and Simon Dixon, ‘Approximate Note Transcription for the Improved Identification of Difficult Chords’, in Proceedings of the 11th International Society for Music Information Retrieval Conference, ed. J. Stephen Downie and Remco C. Veltkamp (Utrecht, the Netherlands, 2010), pp. 135–40, http://ismir2010.ismir.net/proceedings/ismir2010-25.pdf.
To the extent possible under law, the DDMAL has waived all copyright and related or neighbouring rights to the McGill Billboard annotations. This work is published from Canada.