Overview of Research at IITB Computational studies on Hindustani music
CompMusic Workshop, Chennai 2013
Preeti Rao
Department of Electrical Engineering I.I.T. Bombay
1 Some goals • Automatic “tagging” of audio by genre, style, raga, tala and other discovered descriptors and relationships
• Automatic creation of “navigation layer” for concert audio
• Facilitating search for musically relevant objects such as melodic and rhythmic motifs
• Building tools that facilitate musicological research on performance practices
Common to all the above: Need for a music representation (aka features) and similarity measure (for classification)
2
Music CD cover information… (YouTube has even less!)
3 Hindustani music descriptors/tags
• Artiste (instrument), accompanists
• Genre (dhrupad, khyal, tarana…), sub-genre (gharana)
• Concert structure and sections with timing – Bada khyal, Chhota khyal • Bandish: alap, vistaar, taan
• Raga, Tala, Laya of major sections
• Composition (bandish, identified by mukhda)
The question Can the descriptions be obtained by audio content analysis and possibly enhanced with contextual semantic information?
4 Kishori Amonkar Deshkar: bada khyal (vistaar, taan)
Alap (slow tempo)
Alap (medium tempo)
Taan
5 Kishori Amonkar Deshkar: bada khyal taan, chhota khyal taan
Taan (madhylaya)
Taan (drut laya)
6 Kishori Amonkar Deshkar
vistaar alap taan chhota khyal (bada khyal) (bada khyal) 7 Uday Bhawalkar (dhrupad) Yaman
Alap
Jod (alap)
Jhala (alap)
8 Uday Bhawalkar (dhrupad) Yaman
alap jod jhala
9 Kishori Amonkar: Deshkar, Gaud-Sarang Raga Deshkar
Raga Gaud Sarang
10 original resynthesized Raga Deshkar
Raga Gaud Sarang
11 Raga Deshkar
P P P P G G G G G G R S R R S R S S R S S D R S
Raga Gaud Sarang
R G S S S S S S S R D D S R D D D 12 Raga characteristics from pitch distribution
Pt. Vidhyadhar Vyas Marwa and Puriya (share the same swaras)
13 Melodic motif (mukhda) detection
Kishori Amonkar, Deshkar, Tintal Bandish: Piya Jaag 14 Phrase duration, dependence on tempo
15 Within-concert variability of motif
16 Within-concert variability across concerts Intra-phrase class distance distribution
Kafi
17 Ontology for Indian music
Learning metadata (textual) from forums by using NLP techniques to learn relationships between entities.
• Augment audio-based music ontology for Indian music with information extracted from online music forums to achieve superior retrieval systems for Indian music.
Example: Get songs with phrase ‘NDNP’ and sung by a disciple of D.K. Pattammal
• Challenges: text sources are unstructured, ungrammatical…
18 Summary
• High-level musical attributes (melody, rhythm) can be derived from low-level acoustic parameters such as pitch and onsets extracted by audio signal processing.
• The audio signal processing is challenging due to the mixture of several instruments, strong diversity in the characteristics and the highly time-varying nature.
• Melody and rhythm representations that are musicologically informed can be useful in the description of music recordings.
• Knowledge can be very helpful. Creating these from available sources is a challenge.
19 Coming up…
• Vedhas Pandit, Kaustuv : Characterization of melodic motifs
• Vinutha T. P.: Rhythmic structure based segmentation
• Joe Cheri Ross: Ontology for Indian Music: An Approach for ontology learning from online music forums
• Amruta J. Vidwans and Prateek Verma: Melodic style detection in Hindustani music
20 Thank you
Department of Electrical Engineering , IIT Bombay 21