CORPUS

corpus_logo

The COFLA research corpus

Research on computational analysis of famenco music, despite it being a relatively new field, can provide powerful tools for the discovery and diffusion of flamenco music. In the scope of the COFLA project, we compiled a research corpus containing more than 1800 songs which serves as a pool for the creation of datasets for specific music information retrieval tasks.

Below, we provide a short description of the corpus and various annotated datasets. All manual and automatic annotations as well as meta-data are available for download. Please note, that due to copyright restrictions, the full audio tracks are not publicly available. All recordings contained in the corpus are taken from commercially available flamenco anthologies. If you are interested in obtaining the audio data for research purposes, please send us a request.

Publications

This work has been accepted for publication in the ACM Journal of Computation and Cultural heritage and is currently available in arXiv.
N. Kroher, J. M. Díaz-Báñez, J. Mora and E. Gómez (2015): Corpus COFLA: A research corpus for the Computational study of Flamenco Music. arXiv:1510.04029 [cs.SD cs.IR].

Conditions of use

The provided datasets are offered free of charge for internal non-commercial use. We do not grant any rights for redistribution or modification. All data collections were gathered by the COFLA team.
© COFLA 2015. All rights reserved.


Description

corpusCOFLA

The corpusCOFLA is a collection of more than 1500 flamenco recordings which are representative of what is considered classical flamenco. All contained tracks are taken from 12 commercially available flamenco anthologies in order to minimize a possible bias towards geographic location, singer or record label. We provide the editorial meta-information together with the musicBrainz IDs for all tracks as well as the anthologies as XML documents.

Content:

  • corpus meta data (619KB): XML file containing editorial meta-information for all tracks: source (anthology, CD number, track number), artist, title, style and musicBrainzID.
  • anthology meta data (3KB): XML file containing editorial meta-information for all anthologies comprising the corpus: name, record label, year edition, year re-edition, number of CDs

Versions:
Version 0 (released Feb. 6th, 2016): initial release.
Version 1 (released Nov 23rd, 2017):

  • the anthology “Antología del Cante Flamenco. Flamencología.” is no longer commercially available and has been removed from the corpus
  • in the corpus meta-data, a field “style_annotated” has been added, which contains unified styles annotations
  • singer names have been assigned unique identifiers

cante2midi

The cante2midi dataset contains 20 tracks taken from the corpus and includes a large variety of styles and complexity with respect to melodic ornamentation. We provide note-level transcriptions of the singing voice melody in a MIDI-like format, where each note is defined by onset time, duration and a quantized MIDI pitch. In addition, we provide a number of low-level descriptors and the fundamental frequency corresponding to the predominant melody for each track. The meta-information includes editoral meta-data and the musicBrainz IDs.

Content:

  • README (5KB): Text file containing detailed descriptions of manual and automatic annotations.
  • meta-data (10KB): XML file containing meta-information: Source (anthology name, CD no. and track no.) and editorial meta-data (artist name, title, style and musicBrainzID).
  • manual transcriptions (82KB): MIDI (.mid) and text files (.notes) containing manual note-level transcriptions of the singing voice.
  • automatic transcriptions (75KB): Text files (.notes) and MIDI files (.mid) containing automatic note-level transcriptions of the singing voice.
  • Bark band energies (39.9MB): Text files (.csv) containing the frame-wise extracted bark band energies.
  • predominant melody (6.2MB): Text files (.csv) containing the frame-wise extracted predominant melody.
  • low-level descriptors (7.9MB)Text files (.csv) containing a set of frame-wise extracted low-level features.
  • MFCCs (17.8MB): Text files (.csv) containing the frame-wise extracted mel-frequency cepstral coefficients (MFCCs).
  • Magnitude spectrum (709.1MB, optional): Text files (.csv) containing the frame-wise extracted magnitudes of the discrete fourier transform (DFT)

cante100

The cante100 dataset contains 100 tracks taken from the corpus. We defined 10 style families of which 10 tracks each are included. Apart from the style family, we manually annotated the sections of the track in which the vocals are present. In addition, we provide a number of low-level descriptors and the fundamental frequency corresponding to the predominant melody for each track. The meta-information includes editoral meta-data and the musicBrainz ID.

Content:

  • README (5KB): Text file containing detailed descriptions of manual and automatic annotations.
  • meta-data (59KB): XML file containing meta-information: Source (anthology name, CD no. and track no.), editorial meta-data (artist name, title, style, musicBrainzID) and the manually annotated style family.
  • vocal sections (8.9MB): Text file (.csv) containing frame-wise vocal section annotations.
  • automatic transcriptions (375KB): Text files (.notes) and MIDI files (.mid) containing automatic note-level transcriptions of the singing voice.
  • Bark band energies (216.6MB): Text files (.csv) containing the frame-wise extracted bark band energies.
  • predominant melody (33.5MB): Text files (.csv) containing the frame-wise extracted predominant melody.
  • low-level descriptors (42.9MB): Text files (.csv) containing a set of frame-wise extracted low-level features.
  • MFCCs (97.1MB): Text files (.csv) containing the frame-wise extracted mel-frequency cepstral coefficients (MFCCs).
  • Magnitude spectrum (3.85GB): Text files (.csv) containing the frame-wise extracted magnitudes of the discrete fourier transform (DFT).

canteFAN

The canteFAN dataset contains 10 recordings of the fandango style taken from the corpus. We manually annotated repeated melodic patterns on a track-leve. In addition, we provide a number of low-level descriptors and the fundamental frequency corresponding to the predominant melody for each track. The meta-information includes editoral meta-data and the musicBrainz ID.

Content:

  • README (5KB): Text file containing detailed descriptions of manual and automatic annotations.
  • meta-data (5KB): XML file containing meta-information: Source (anthology name, CD no. and track no.) and editorial meta-data (artist name, title, style, musicBrainzID).
  • pattern annotations (6KB): text (.txt) files containing the manually annotated repeated melodic patterns.
  • automatic transcriptions (28KB): Text files (.notes) and MIDI files (.mid) containing automatic note-level transcriptions of the singing voice.
  • Bark band energies (16.6MB): Text files (.csv) containing the frame-wise extracted bark band energies.
  • predominant melody (2.6MB): Text files (.csv) containing the frame-wise extracted predominant melody.
  • low-level descriptors (3.3MB): Text files (.csv) containing a set of frame-wise extracted low-level features.
  • MFCCs (7.4MB): Text files (.csv) containing the frame-wise extracted mel-frequency cepstral coefficients (MFCCs).
  • Magnitude spectrum (290.1MB): Text files (.csv) containing the frame-wise extracted magnitudes of the discrete fourier transform (DFT).

Download

To download the manual annotations and meta-data, please fill in the download form.
If you are interested in obtaining the audio files, please send us a request.