Week 3 — Tune It Up
Hello world,
We are Fidan Samet, Oğuz Bakır and Adnan Fidan. In the scope of the Fundamentals of Machine Learning course project, we are working on prediction and style transfer of song release years. We will be writing blogs about our progress throughout the project and this is the third one of our blog series. The datasets considered in the related works of our tasks will be covered in this post. So let’s get started!
Previously on Tune It Up…
Last week, we talked about the methods and approaches considered in the related works of our tasks. You can find last week’s blog here. This week, we will examine the datasets we plan to use in our project.
Datasets
1. Million Song Dataset¹
In this dataset, the authors perform song year prediction as a case study. Since they perform song year prediction, they need lots of music for feature extraction from a large interval of years. Later, these features are used for classification models. Below is the table of data entry fields in this dataset.
This dataset, as the name implies, includes 1.000.000 song entries without any real audio files. Instead of audio files we can use segments_timbre field for timbre features. Timbre features can be used for both classification and style transfer of the songs. In addition to timbre features, we can use year and genre fields to classify songs according to decades and also according to decades and genres. Since genre field is not populated on MSD, we can create their genres using MusicBrainz² tags with script³. Below is the table of genre distribution on MusicBrainz.
Using MusicBrainz tags, we can achieve distribution of genres as seen in above table. It can be seen that most of the entries are in “classic pop and rock”, “folk” and “rock and indie” genres. Below are the graphs of MSD tracks which have release year information.
MSD has a right skewed distribution in release year distribution graph which makes highly unbalanced distribution over entries. Because of these unbalanced entries, we aim to use genres with 1965-present time interval. In addition to this limitation, we aim to use time intervals with 10 year jumps as decades. Despite these time interval and genre limitations, we can perform song classification and style transfer.
2. FMA: A Dataset For Music Analysis⁴
FMA contains real audio files for each entry but it is highly unbalanced on release dates which can be seen in below figure. In addition to this unbalance, release years start from 1990 and they yield too narrow year interval for our task. Below is the graph of FMA tracks which have release year information.
Since there is high unbalance on release years, we are not planning to use this dataset on future works for now.
That is all for this week. Thank you for reading and we hope to see you next week!
References
[1] Teixeira, M., & Rodrguez, M. M0444 Project One: Release Year Prediction for Songs.
[2] https://musicbrainz.org/
[3] http://millionsongdataset.com/blog/11-2-28-deriving-genre-dataset/
[4] Defferrard, M., Benzi, K., Vandergheynst, P., & Bresson, X. (2016). Fma: A dataset for music analysis. arXiv preprint arXiv:1612.01840.
