Silasimo/NUS-48E-Segmented

This dataset contains a segmented version of the NUS-48E dataset (original paper, original data). The original dataset was shared with the agreement that the dataset will be used for research purposes only, thus we will impose the same requirement for this segmented version.

The NUS-48E segments were created as a singing-oriented forced alignment benchmark in our SynthGT project. Segmentation points were chosen manually to create phrase-level excerpts. The word-level separator sp has been removed, and sil has been replaced with SP (silent pause). This was done to share the same vocabulary with, and have comparable audio length to, our SynthGT dataset. We share the audio files in the wav folder, along with the full annotations in TextGrid format in the TextGrid folder, and the phoneme sequences in the lab folder. We use the wav and lab pairs to predict phoneme durations, and compare the result to the respecive TextGrid to evaluate forced alignment performance. Our models which have been evaluated with the NUS-48E-Segmented benchmark can be found in our model collection.

If you use this dataset in your own research, please cite our work:

bibtex
@unpublished{Antonisen2026SynthGT,
  author = {Silas Antonisen and Iván López-Espejo},
  title  = {Synthesizing Ground Truth Phoneme Boundaries with SynthGT: A Synthetically Generated Solo-Singing Corpus},
  note   = {Submitted to IEEE Transactions on Audio, Speech and Language Processing},
  year   = {2026}
}

As well as the original NUS-48E paper:

bibtex
@inproceedings{NUS48E,
  title = {{The NUS Sung and Spoken Lyrics Corpus: A Quantitative Comparison of Singing and Speech}},
  booktitle = {Proceedings of the 2013 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)},
  author = {Zhiyan Duan and Haotian Fang and Bo Li and Khe Chai Sim and Ye Wang},
  year = {2013},
  month = {October 29--November 1},
  pages = {1--9},
  address = {Kaohsiung, Taiwan}
}