This repository contains code for data preprocessing for lyrics alignment and transcription tasks. The aim is to reduce the hassle of data preprocessing on DALI v2, MulJam v2, and MultiLang Jamendo v1.1, including downloading, source separation, and segmentation.
Specify where to save the data by setting the environment variable WORKING_PATH
to the root directory of this repository.
export WORKING_PATH=/path/to/save/data
Please have a look at data_util.py
and customize it (datasets, paths, etc.).
Access to DALI dataset needs to be granted: https://zenodo.org/records/3576083
To download MulJam and Jamendo, please run the following command and specify the dataset (MulJam
or Jamendo
):
python download.py --dataset MulJam
All the audio files will be saved in mix_dir
.
To separate the vocals from the music, we use the pretrained Demucs model.
By default, it applies on all the audio files in mix_dir
. To specify a subset, please provide the --ids_file
argument:
python separate.py [--ids_file ./input/dummy/dali_dummy_ids.npy]
All the source-separated files will be saved in sepa_dir
.
For segmentation, a metadata file is required to specify the start and end time of each lyric line.
It needs to have the following information for each line (see ./input/dummy/*.meta
for example)
seg_id,song_file,line_start,line_end
Also specify the --split
argument to indicate the split of the dataset (train, valid, or test),
because for training set we distribute files by their prefixes to improve file system performance.
python segment.py --meta_file ./input/preconstructed-split/jamendo_line.meta --split test
All the segmented files will be saved in seg_dir
, and the utterances are organized as the following structure:
$seg_dir
├── train
├── 00/000_0.wav
├── 42/420_1.wav
└── ...
├── valid
├── 126_43.wav
├── 731_9.wav
└── ...
└── test
├── 999_0.wav
├── 999_1.wav
└── ...
Data splits for the latest version datasets are provided in ./input/preconstructed-split/
.
Please read the README.md in that directory for more details.