4MuLA - A Multitask, Multimodal, and Multilingual Dataset of Music Lyrics and Audio Features

• Angelo Cesar Mendes da Silva USP
• Diego Furtado Silva UFSCar
• Ricardo Marcondes Marcacini USP

Resumo

We present a new benchmark dataset of songs with structured information to be applied in various machine learning tasks. The data comes from a platform focused on lyrics information, but contain several other annotations provided by their users. Our dataset, called 4MuLA (Multitask, Multimodal, and Multilingual Music Lyrics and Audio features dataset), includes features extracted from 96,458 songs distributed by 15,310 artists in 76 genres. In particular, our dataset contains latin music genres that are often overlooked in other benchmark datasets. For each track, we make available various acoustic features, extracted tags, and lyrics in English, Portuguese, or Spanish. With these features, researchers can use our dataset for, at least, lyrics-, audio- or multimodal-based genre classification, music and artist similarity, and popularity regression. Moreover, we can perform cross- or multilingual text analysis on lyrics, such as discourse analysis or measuring the differences between emotion transmitted by audio and lyrics.

Palavras-chave: music dataset, multimodal musical dataset, latin musical dataset