DAP-Lab / dhrupad-bandish-segmentation

Dhrupad Vocal Bandish Segmentation

This repository contains:

• annotations of metric cycle (tala) boundaries and of surface tempo based structural segments of the bandish portions of a Dhrupad vocal concert dataset
• codes for automatic surface tempo multiple estimation
  

It is linked to the following publications:

Rohit M. A., Vinutha T. P., Preeti Rao. “Structural Segmentation of Dhrupad Vocal Bandish Audio 
based on Tempo”, 21st International Society for Music Information Retrieval Conference,
Montréal, Canada, 2020

Rohit M. A., Preeti Rao. “Structure and Automatic Segmentation of Dhrupad 
Vocal Bandish Audio”, Unpublished technical report, arXiv:2008.00756 [eess.AS], 2020

The annotations were created manually by one of the authors in consultation with a musician. Trained models are made available to obtain predictions on any test audio. Training scripts are also provided to reproduce the results reported in the paper.

Contents

• annotations - annotations for all the audios used in the paper
  

• codes - scripts to use the trained models to obtain surface tempo multiple estimates on a test audio, as well as to reproduce the cross-validation results
  

• audio_samples - mixture and source separated audio clips corresponding to sections (a) and (b) of Figure 2 in the paper

• docs - PDF files of the submitted paper and the technical report

More details on the annotation format and running the codes can be found in the respective folders.

Audio dataset

The sources for all the audios used in the work are listed in the file dataset_sources.pdf. Some are available on YouTube, while others are from the CompMusic Dunya ¹ collection and can be obtained through the Dunya API ² using the provided MusicBrainz IDs ³.

---

¹https://dunya.compmusic.upf.edu/
²https://dunya.compmusic.upf.edu/developers/
³https://musicbrainz.org/doc/MusicBrainz_Identifier