The Latin Music Database A Database for Automatic Music Genre Classification Carlos N. Silla Jr., Celso A. A. Kaestner, Alessandro L. Koerich 11 th Brazilian Symposium on Computer Music (SBCM2007) São Paulo Setembro/2007
Outline Introduction Other Databases Main Characteristics of the LM Database Experiments with the LM Database Limitations Concluding Remarks Future Work
Introduction Problem Statement It is very difficult to assign a genre to a music based only on the auditive human perception It implies in a previous knowledge about the genre Lack of ground-truth in the area Most of databases normally contains few musical recordings, sometimes only excerpts of the full music and also a small number of recordings per class
Introduction Motivation To build a database for research of machine learning algorithms To build a database where the music pieces are labeled by specialists Minimize the subjectiveness of assigning a genre to a music piece
Introduction Goal To build a clean database to carry out experiences on automatic music genre classification using machine learning algorithms. To build a database that allows reproducing experiments No duplicity of music pieces Make easy the input of new music pieces and genres
Other Databases CODAICH 20.894 music pieces in MP3 format from 1.941 artists. GTZAN Database 1.000 music pieces from 10 genres (100/genre) Homburg 1.886 music pieces from 9 genres
Main Characteristics of the LM Database How the genres were assigned to the music pieces? Human Inspection Based on the perspective of the human perception on how the music is danced. By professional dance teachers with over ten years of experience in teaching ballroom and Brazilian cultural dances.
Main Characteristics of the LM Database First Stage Dance teachers make a selection of the musical recordings that they judged representative of a specific genre, according to how that musical pieces are danced. Second Stage Each selected music piece was verified to avoid mistakes that were expected to happen due to the stress produced by manually listening and labeling each one of the pieces.
Main Characteristics of the LM Database About 300 musical recordings were classified by month, and the total duration of the development of the Latin Music Database took a year. The Latin Music Database: 3.160 music pieces in MP3 format 10 different musical genres 543 artists
Main Characteristics of the LM Database Musical Genres and the number of samples Tango (404) Salsa (303) Forró (315) Axé (304) Bachata (308) Bolero (302) Merengue (307) Gaúcha (306) Sertaneja (310) Pagode (301) At least 300 samples per genre
Main Characteristics of the LM Database The procedure to insert a music piece in the database: 1. Assign a genre (specialist) 2. Inspection and correction of the ID3 tag (artist and title). 3. Enrollment of the music piece in the database DIRECTORY_GENRE\ARTIST-TITLE-ALBUM- TRACK.MP3
Main Characteristics of the LM Database Other approaches CD Collections Artist Profile In the case of Latin Music, such approaches have some drawbacks.
Main Characteristics of the LM Database Example: In the 4 CD collection Los 100 Mayores Exitos De La Musica Salsa only half (50 out of 100) of the music pieces can be considered as Salsa! Only 400 out of more than 500 Carlos Gardel compositions are really Tango! While building the database, we have found that in average, one to three music pieces are in conflict with the artist profile
Experiments with the LM Database
Experiments with the LM Database C. N. Silla Jr., C. A. A. Kaestner & A. L. Koerich. Automatic Music Genre Classification Using Ensemble of Classifiers. IEEE International Conference on Systems, Man and Cybernetics (SMC2007), Montreal, Canada, to appear, October 2007.
Concluding Remarks A novel approach to assign genres to music pieces Labeled by specialists Based on how a music piece is danced. Extends the auditive human perception Maybe it is the first thematic database
Limitations The raw data (MP3 files) is not available. 30-dimensional feature vectors generated from the Marsyas framework 30 are publicly available at www.ppgia.pucpr.br/~alekoe
Future Work Make it available at the On-demand Metadata Extraction Network (OMEN project). A tool that overcomes the copyright limitation and make databases widely available to the MIR community Inclusion of new musical genres to the current database Introduce an hierarchy of genres. Example: forró will be the main genre of subgenres xote, xaxado and baião.
Questions?