README: Automatic Sung-Lyrics Data Annotation

Created by: Chitralekha Gupta

Affiliation: NUS, Singapore

Last edited on: 12th June 2018

Last edited by: Chitralekha Gupta

This is the code base for automatically obtaining aligned lyrics for solo-singing in Smule's Sing! karaoke DAMP dataset. It provides a cleaner annotated dataset.

Please refer to the following paper for details:

Gupta, C., Tong R., Li, H. and Wang, Y., 2018, September, "SEMI-SUPERVISED LYRICS AND SOLO-SINGING ALIGNMENT". Accepted for ISMIR 2018, Paris.(http://ismir2018.ircam.fr/doc/pdfs/30_Paper.pdf)

Contents

This consists of the following:

• Audio folder (folder: audio), where an example audio (.m4a) from DAMP dataset is provided
• Converted .m4a to .wav are in the folder: wavfiles
• All ~10 seconds segments are in the folder: wavsegments_initial
• Cleaned up subset of segments are in the folder wavsegments_final, the corresponding automatically obtained lyrics annotations are in "resultant.txt"
• fulloutput.txt contains all segment names, lyric window transcript, ASR transcript, and %correct
• Lyrics folder (folder: lyrics), where lyrics are extracted from Smule Sing! website
• 'perfs20.csv' is a meta-data file from Kruspe's dataset, given here: http://www.music-ir.org/mirex/wiki/2017:Automatic_Lyrics-to-Audio_Alignment
• Rest are python scripts and other dependency files.

Dependencies

• This program is designed for monophonic (without background music) audio files.

How to run?

• The python scripts main.py is the main file.
• Start with main.py, and follow the instructions in the file header.
• The script currently runs for one example file present in the folder "audio"
• You can add more files from the DAMP dataset to the "audio" folder

Contact

• Chitralekha Gupta: chitralekha[at]u[dot]nus[dot]edu
• Haizhou Li: haizhou[dot]li[at]nus[dot]edu[dot]sg
• Ye Wang: wangye[at]comp[dot]nus[dot]edu[dot]sg