Indian Art Music Raga Recognition Dataset

Introduction

Rāga recognition datasets comprise two sizable datasets, one for each music tradition, Carnatic music dataset (CMD) and, Hindustani music dataset (HMD).These datasets comprise full length audio recordings and their associated rāga labels. These two datasets can be used to develop and evaluate approaches for performing automatic rāga recognition in Indian art music. To the best of our knowledge, these are the largest
and the most comprehensive (in terms of the available metadata) datasets ever used for studying this task. For more information about the dataset we refer to Chapter 3 of this thesis. 

These datasets are derived from the CompMusic corpora of Indian art music, for which each recording is associated with a MBID. With the MBID other information can be obtained using the Dunya API. 

Please cite the following publications if you use the material shared here in your research work.

[1]. Gulati, S., Serrà, J., Ganguli, K. K., ¸Sentürk, S., & Serra, X. (2016). Time-delayed melody surfaces for raga recognition. In Proceedings of the 17th International Society for Music Information Retrieval Conference (ISMIR), pp. 751–757. New York, USA.

[Postprint PDF@MTG] [Bibtex] [Resources]

[2]. Gulati, S., Serrà, J., Ishwar, V., ¸Sentürk, S., & Serra, X. (2016). Phrase-based raga recognition using vector space modeling. In Proceedings of the 41st IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 66–70. Shanghai, China.

[Postprint PDF@MTG] [Bibtex] [Resources]

Carnatic music rāga recognition dataset

CMD comprises 124 hours of audio recordings and editorial metadata that includes carefully curated and verified rāga labels. It contains 480 recordings belonging to 40 rāgas with 12 recordings per rāga. The details of this dataset in terms of the number of asscociated musical entities and the relationships between them is provided below. 

(click to enlarge images)

 

 

The CMD raga recognition dataset is version controlled and can be accessed from here. It also contains extracted audio features (pitch + tonic), which are also made publicly available here. To know how the features are extracted visit the companion page for the publication.

Hindustani music rāga recognition dataset

HMD comprises 116 hours of audio recordings and editorial metadata that includes carefully curated and verified rāga labels. It contains 300 recordings belonging to 30 rāgas with 10 recordings per rāga. 

(click to enlarge images)

 

 

 

The HMD raga recognition dataset is version controlled and can be accessed from here. It also contains extracted audio features (pitch + tonic), which are also made publicly available here. To know how the features are extracted visit the companion page for the publication.

The audio files corresponding to these datsets are made available on request for only research purposes. To obtain the files fill this FORM.

After downloading the above mentioned audio features and audio recordings place them according to this directory structure.

Annotation Format

We provide both tsv files and json files that contain information about each audio recording in terms of its mbid, the path of the audio/feature files and the associated rāga identifier. Each rāga is assigned a unique identifier by Dunya, which is similar to the mbid in terms of purpose. We also provide a mapping of the rāga id to its transliterated name.

Contact