ohollo/lmd_chords

Dataset Card - Chords from the Lakh MIDI Dataset

Dataset Summary

The Chords from the Lakh MIDI Dataset (LMD) is a collection of 31032 chord sequences extracted from selected MIDI files of the Lakh MIDI Dataset extracted using the Python library chord-extractor, which has the ability to take MIDI, MP3, WAV and other sound files in bulk, and extract chords using the Chordino method.

The music files chosen are those that the LMD Dataset curator was able to best match to songs in the Million Song Dataset. In doing so we are able to include here the song title, artist and other metadata associated with the chord sequences.

Dataset Structure

Data Instances

A typical data point comprises the chord sequences (with when they occur in the song), IDs so that one can source the metadata or actual MIDI files themselves, along with some information about the original piece of music that the MIDI is based on.

An example from the dataset looks as follows:

{
  'title': "Movin' Out (Anthony's Song)",
  'artist': 'Billy Joel',
  'track_id': 'TRNCSKU128F4265639',
  'year': 1977,
  'song_id': 'SOIJWHG12A8C134063',
  'chords': {
      'symbol': ['N', Dm', 'Bb', 'C','Eaug', 'Am', 'Dm', 'Gm7', 'C', 'Eaug', 'Am', ...],
      'timestamp': [0.371519274, 0.4643990993499756, 1.8575963973999023, 3.7151927947998047, 4.551111221313477, ...]
  }
}

Data Fields

• title: The title of the song whose chords have been interpreted.
• artist: Artist/performer of the song.
• track_id: The track ID as used by Echo Nest, a source of data for the Million Song Dataset. This is also important as it is used in LMD's "LMD-matched" dataset as the directory names containing the MIDIs corresponding to a particular song from the MSD. Therefore use this to navigate those files.
• year: Year that the song was first released (see considerations) if provided. If not provided by MSD, this is denoted 0. Of 31032 total records, 16832 have years.
• song_id: "Song ID" as provided by the Million Song Dataset.
• chords: Two lists of same length describing chord progressions in song
  • symbol: List of chord symbols in the order as they appear in the song. The list is bookended by Chordino with two dummy 'N' symbols to assist with when the track starts and ends.
  • timestamp: Timestamps (in seconds) at which the chords appear in the song. The index of the timestamp, corresponds to the chord symbol in the same index as "symbol".

Data Splits

The data currently has one split - "train".

Dataset Creation

Curation Rationale

The LMD Chords are created to provide a chord sequence dataset of reasonable size such that potentially good insights can be made into harmonic progression. Enhanced analysis is also possible with the timestamps of these chords provided. The selection of files matched with MSD should give a decent representation of different genres, years, artists etc.

With the lack of accessible large publicly available datasets with the actual original songs, for example in MP3 format, MIDIs are seen as a viable alternative, as they should relay the notes played in the song, despite not being timbrally the same as the original.

Furthermore, the IDs of the Million Song Dataset provided give the user the opportunity to map the chords to the wealth of audio analysis held in the full version of the MSD. These not only include an expanded version of the metadata provided here, but features such as tempo, key, loudness and much more.

Source Data

Initial Data Collection and Selection

To source the music files used for chord extraction, three files are needed for download:

• The set of MIDI files that have been matched to the Million Song Dataset, called "LMD-matched" available on the LMD website.
• On the same page, are the Match Scores, that show the match confidence score between each MIDI file and the associated MSD record. This JSON file contains a map between MSD/Echo Nest track IDs and the associated MIDI filenames in LMD and their score.
• The Summary File of the whole Million Song Dataset, available on this page. Note this is a 300MB HDF5 with just the metadata. The whole MSD (which is 280GB) is not needed to create this dataset.

The LMD Chords dataset only uses one MIDI per song. The Match Scores file was used to see the best matching MIDI to a particular song from MSD, and that file was then included here.

The name of a directory holding a particular MIDI is the same as the track ID for that song used by MSD, and therefore was used to map a MIDI file to the MSD and source its metadata.

Chord Extraction Process

The selected MIDI files are then passed to the chord-extractor v0.1.2 (full disclosure: this library is also written by the creator of this dataset). The Chordino method is utilised, with an override on the spectral roll on parameter set to 1 as below (note, the extract_many method automatically converts MIDIs to WAV prior to extraction).

midi_paths = [...] # List of strings denoting paths to LMD MIDIs matched to MSD
extractor = Chordino(rollon=1)
res = extractor.extract_many(midi_paths) # Result provided to symbol and timestamp fields in dataset

Please refer to this website to read more information about Chordino and the NNLS Chroma technique it uses.

Considerations for Using the Data

Limitations of the Chord Extractions

The chords listed here are interpretations made by the Chordino method, an algorithm applied to sound files, and as such will likely deviate from official transcriptions and scores. The chord symbols effectively provide root notes, differentiations between major and minor chords, major and dominant 7ths, as well as providing slash chords, and some others, the chords types available are not exhaustive. It is not able for example to identify complex jazz chords such as Fmaj7#11.

It should also be noted that the MIDIs were converted to WAV sound files prior to extraction (as this was the format needed by chord-extractor). In that way some information may be lost, compared to directly gleaning chords from reading the note information available in the MIDI file.

Nevertheless, the extractions are of sufficient quality to give a good semblance of the "ground truth", particularly for harmonically non-complex songs, which form the majority. The user is encouraged to compare some of the extractions here to their expectations.

Accuracy of the Metadata

It should be noted that though the Million Song Dataset has sourced its metadata from The Echo Nest, it is not immune from the occasional typo, though this is observed to be very rare with regard to song name and artist. Other metadata in MSD can be open to interpretation (e.g. genre). The user should familiarise themselves with how MSD has populated a field, if heavily relying on its data. The year field as relayed in this dataset is a good example of this. It is only present for approximately half of the data points, and though it is assumed to show the original year of the song's release, it should be verified that this is always the case.

Legal and Ethical Considerations

This dataset contains interpretations of chord progressions from musical works. While chord progressions themselves are not typically protected under copyright law (and the author is not aware of any infringement here), the original works they are derived from remain copyrighted. Users are encouraged to respect the rights of the original artists when using this dataset.

This dataset is provided for research and educational purposes. Users should ensure their use of the dataset complies with applicable copyright laws and licensing agreements in their jurisdiction.

Additional Information

Dataset Curators

The dataset was collated by Oliver Holloway.

Licensing Information

The dataset is licensed under CC BY-SA 4.0. Please review the license for detailed usage terms.

Citation Information

If using this dataset, please cite it by getting the latest Bibtex from the "Cite this dataset" button on this page. It should resemble the following:

@misc {oliver_holloway_2025,
    author       = { {Oliver Holloway} },
    title        = { lmd_chords (Revision 4d6815c) },
    year         = 2025,
    url          = { https://huggingface.co/datasets/ohollo/lmd_chords },
    doi          = { 10.57967/hf/4219 },
    publisher    = { Hugging Face }
}

It would be great to mention the chord-extractor Python library as well.

Furthermore, this dataset is derived from the Lakh MIDI Dataset so please cite the below:

Colin Raffel. "Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-MIDI Alignment and Matching". PhD Thesis, 2016.

If you use the metadata of this dataset, it is from the Million Song Dataset, so please reference this paper:

Thierry Bertin-Mahieux, Daniel P. W. Ellis, Brian Whitman, and Paul Lamere. "The Million Song Dataset". In Proceedings of the 12th International Society for Music Information Retrieval Conference, pages 591–596, 2011

Finally, the chords have been derived using the Chordino method so please cite:

M. Mauch and S. Dixon, “Approximate Note Transcription for the Improved Identification of Difficult Chords,” in Proceedings of the 11th International Society for Music Information Retrieval Conference (ISMIR 2010), 2010.