Welcome to the companion web site of our paper “On strategies to exploit dependencies between singing voice alignment and separation”. This page provides additional material and resources related to the paper.

Datasets

We release here three types of additionnal data related to the MUSDB18 database:

1. Graphemic annotations aligned with songs of the MUSDB18 dataset. The annotations are provided in the form of LAB files, which can be downloaded here.

2. Phonetic annotations aligned with songs of the MUSDB18 dataset. The annotations are provided in the form of LAB files, which can be downloaded here.

3. Posteriorgrams of the MUSDB18 dataset songs. The posteriorgrams are provided by the ADAGIO system in the form of NPY files, which can be downloaded here (78G needed).

You may also be interested in the raw text lyrics of the MUSDB18 dataset, which can be downloaded here.

Split in training, validation and test sets

We provide the split of the MUSDB18 dataset in training, validation and test sets used in our experiments. The split is provided in the form of TXT files, which can be downloaded here.