3rd CompMusic Workshop

Date: December 13th, 2013
Venue: CS 25, Dept of Computer Science and EngineeringIndian Institute of Technology Madras, Chennai (India)
Scientific Committee: Xavier Serra (UPF), Preeti Rao (IITB), Hema Murthy (IITM)

This workshop aims to give an overview of the research being done within the CompMusic project related to Hindustani and Carnatic music. The workshop is followed, on the next day, by a special seminar dedicated to the discussion of the tools that can be used for the exploration and appreciation of Carnatic music, and on December 15th there will be a Lecture Demonstration by Padma Vibhushan Sangita Kalanidhi Mrudangam Legend Dr. Umayalpuram K Sivaraman. Entrance to all the events is free but please confirm your attendance to Prof. Hema Murthy


09:00h Xavier Serra: “CompMusic research progress: halfway results

09:30h Gopala Krishna Koduri: “Knowledge-based representations for Carnatic music

09:50h Sankalp Gulati: “Representation and segmentation of melodies in Indian Art Music

10:10h Ajay Srinivasamurthy: “Rhythm analysis and tāḷa tracking in Carnatic music

10:30h Coffee break

11:00h Alastair Porter: “Dunya: A system for browsing audio music collections

11:20h Preeti Rao: “Overview of research done at IIT-Bombay

11:40h Vedhas Pandit and Kaustuv Kanti Ganguli: “Characterization of melodic motifs

12:00h T. P. Vinutha: “Rhythmic structure based segmentation

12:20h Lunch

14:00h Joe Cheri Ross: “Ontology for Indian Music: An Approach for ontology learning from online music forums

14:20h Amruta J. Vidwans and Prateek Verma: “Melodic style detection in Hindustani music

14:40h Hema Murthy: “Overview of research done at IIT-Madras

15:00h Shrey Dutta: “Motif spotting in Carnatic music

15:20h Coffee break

15:50h P. Sarala and Akshay Anantapadmanabhan: “Cent filterbanks and their relevance to Carnatic music

16:10h Rajeev Rajan: “Group delay based melody extraction for Indian music

17:00h End of workshop


Xavier Serra: “CompMusic research progress: halfway results” [slides, video]
In CompMusic we work on computational approaches to analyze and describe music recordings by emphasising the use of domain knowledge of five particular music traditions: Arab-Andalusian (Maghreb), Beijing Opera (China), Turkish-makam (Turkey), Hindustani (North-India) and Carnatic (South-India). Our focus is in describing the melodic and rhythmic characteristic of these musics with the goal to better understand and explore their musical richness and personality. The project started in July 2011 and will continue until June 2016, thus we are at the half way point. An initial effort was the compilation of audio music collections, plus the accompanying information, to be used as the research corpora. Using these collections we have carried out computational analysis on musically relevant aspects such as: tonic, intonation, melodic motives, and rhythmic patterns. Some of the current results have been already integrated into the a newly developed music web browser, Dunya.

Gopala Krishna Koduri: “Knowledge-based representations for Carnatic music” [slides, video]
Efficient data structuring and information management is crucial for retrieving meaningful relationships between various entities in the music. Especially so, when dealing with varied music traditions. Knowledge representation technologies developed in the semantic web context can be utilized to achieve this goal. Further, it will facilitate a knowledge-guided comparison and integration of different models of the same musical concept. But the current development state of ontologies for music is limited in its scope to address these issues. There is a need to develop ontologies for concepts and relationships coming from music theory. In this work, we discuss the melodic concepts that constitute the rāga, the melodic framework of Indian art music,  specifically in the context of creating ontologies. We present a first version of the rāga ontology, and discuss the challenges posed by modeling semantics of its substructures. We also present an evaluation strategy that relies on Wikipedia to validate our ontology’s coverage of musical concepts related to rāga.

Sankalp Gulati: “Representation and segmentation of melodies in Indian Art Music” [slides, video]
Melody representation and segmentation play a crucial role in discovery of musically relevant melodic motives and their classification. We discuss a preliminary approach for melody representation, which in addition to fundamental frequency also considers instantaneous loudness and spectral envelope of the lead melodic source. We also discuss a segmentation approach that incorporates domain knowledge and facilitate in identifying melodically stable regions. Furthermore, in the case of Hindustani music some of these melodically stable regions can be identified as nyas segments, which in turn enables efficient melodic motif discovery. We discuss a classification based approach for nyas segment detection that uses musically relevant set of features.

Ajay Srinivasamurthy: “Rhythm analysis and tāḷa tracking in Carnatic music” [slides, video]
Tāḷa tracking is the automatic tracking of the components and events related to the progression of a tāḷa and can provide listeners with a better understanding of the temporal structure of the music piece. Tāḷa tracking from audio music recordings involves an estimation of several useful rhythm descriptors. We present some of the challenges we have faced and discuss the approaches we have explored for estimating the components of the tāḷa in the context of Carnatic music. In specific, we focus on akshara pulse and sama tracking. We also present a rhythm annotated Carnatic music dataset that can be useful to evaluate the algorithms built for different sub-tasks of tāḷa tracking.

Alastair Porter: “Dunya: A system for browsing audio music collections” [slides, video]
Dunya is a web-based software application that lets users interact with an audio music collection through the use of musical concepts that are derived from a specific musical culture. The application includes a database containing information relevant to particular music collections, such as audio recordings, editorial information, and metadata obtained from various sources. An analysis module extracts features from the audio recordings, which are then used to create musically meaningful relationships between all of the items in the database. The application displays the content of these items, allowing users to navigate through the collection by identifying and showing other information that is related to the currently viewed item, either by showing the relationships between them or by using culturally relevant similarity measures. The basic architecture and the design principles developed are reusable for a variety music collections with different characteristics.

Preeti Rao: “Overview of research done at IIT-Bombay” [slides, video]
At IIT Bombay, we focus on computational methods for Hindustani music with the aim of studying its distinctiveness and developing analyses for salient attributes that can help in the better search and navigation of commercial audios recordings.  Given the structure of the typical Hindustani performance, the segmentation of the concert sections facilitates the application of structure specific melodic and rhythmic analyses. Rhythmic structure and tempo are promising bases for the segmentation of the alap, bada and chhota khayal sections. In the bada khayal section, the detection of the mukhda by repeating motif and other raga characteristic phrases can provide for a rich transcription.  Another research focus is the development of acoustic features for style discrimination in vocal and instrumental music.  Paralleling the work on Carnatic music, we’re also researching learning ontology from online music forums dedicated to Hindustani music. 
Vedhas Pandit and Kaustuv Kanti Ganguli: “Characterization of melodic motifs” [slides, video]
Raga-characteristic phrases or melodic motifs are marked by continuous pitch curves embedded in the detected melody line of a vocal performance. We apply a time-series distance measure to the pitch segments to study the invariance of raga-characteristic motifs within and across concerts of the same raga by several eminent vocalists. We try to interpret the observed limits of the variation in terms of known musicological facts and explore applying these learned limits to the problem of distinguishing raga-characteristic phrases from same-notation phrases of other ragas.
T. P. Vinutha: “Rhythmic structure based segmentation” [slides, video]
Periodicities at various temporal levels are characteristic of the concert segment and can be exploited via tempo and rhythm analyses to achieve the separation of concert sections. We present work on table solos where surface rhythm changes can be detected by periodicity estimates of the stroke onset patterns.
Joe Cheri Ross: “Ontology for Indian Music: An Approach for ontology learning from online music forums” [slides, video]
To enable better information retrieval for music along with representation content based information, music meta data in textual form are also to be a part of music ontology. This work focussing on information extraction from online sources (rasikas.org,wikipedia) has two primary modules: one to extract the relevant information, such as entities, their properties and relationships from each of the aforementioned sources, and another module to map and interlink such information using our ontologies which will result in a multimodal knowledge-base. Natural language processing (NLP) based approach is adopted for information extraction, analysing syntactic patterns present in the content. Entities are identified performing fuzzy matching with the existing knowledge base and relations are extracted through a rule based approach.
Amruta J. Vidwans and Prateek Verma: “Melodic style detection in Hindustani music” [slides, video]
Based on the observations that the melody line alone suffices for listeners to reliably distinguish multiple musical styles, we research computable melodic features to distinguish Hindustani, Carnatic and Turkish music. We also consider distinguishing instrument playing styles based on melodic features.
Hema Murthy: "Overview of research done at IIT Madras" [slides, video]
The work at IIT Madras focuses on computational methods for Carnatic music. Most Rāgas in Carnatic music are based on phraseology and the initial objective was to discover the phrases of a rāga automatically from the recording of songs. A required first step was to identify the tonic, given that its pitch can vary from musician to musician and can also vary from concert to concert. Using signal processing and dictionary learning methods, the tonic is first identified and the melodic contour of a song is normalised with respect to it. The task of motif discovery is difficult owing to the variation in both duration and time warp. As a first step, typical motifs of rāgas are used to find their locations in a relatively long alaapana using a variant of the longest common subsequence algorithm. One approach we have taken to motivic analysis has been based on using cent filterbank based features instead of using the pitch trakcs. These features are then used for different tasks, namely segment recognition, motif recognition and stroke recognition. The segment recognition is used in tandem with applause identification to segment a concert into different items for archival purposes. Most pitch extraction algorithms are based on analazing the magnitude spectrum of the signal. We have also developed new algorithms using the phase spectra, obtaining results comparable to many of state of the art approaches with much less complexity.
Shrey Dutta: “Motif spotting in Carnatic music” [slides, video]
We address the problem of spotting melodic motifs, in an alapana, which are the signature phrases of a raga. This work is based on the previous work where locations of snippets of phrases that exists in the Alapana are identified and then a conventional rough longest common subsequence algorithm is modified to locate the motifs. This work verifies how this approach scales for different ragas.
P. Sarala and Akshay Anantapadmanabhan: “Cent filterbanks and their relevance to Carnatic music” [slides, video]
Carnatic music and Hindustani music are two popular music traditions in India. In both music traditions, the tonic or key is chosen by the performer. All accompanying instruments are tuned to the same tonic. Carnatic music is based on the twelve semitone scales and frequencies of semitones depends on the tonic. The tonic can vary across concerts as well as musicians. In this work, we propose new features that are extracted from cent filterbank energies. Cent filterbank energies are obtained by passing the tonic normalised spectrum through constant Q filters that are placed uniformly on the cent scale. These features are evaluated on different tasks in Carnatic music, namely segmentation, motif recognition, mridangam stroke recognition and other applications.
Rajeev Rajan: “Group delay based melody extraction for Indian music” [slides, video]
We present a novel approach for melody extraction based on the modified group delay function as opposed to conventional magnitude based approaches. In the proposed method, the power spectrum of the music signal is first flattened in order to annihilate system characteristics, while emphasising the source characteristics. The flattened magnitude spectrum is treated as a time domain signal. The modified group delay function of this signal produces peaks at multiples of the pitch period. The first three peaks are used to determine the actual pitch period. To address the effects of pitch doubling or halving, a dynamic programming based approach is used. Dynamic variation of pitch is captured by adaptive windowing in which the window size is determined by fixing a static threshold on autocorrelation of Fourier transform magnitude of frames with a lag. Voicing detection is performed using the normalized harmonic energy. The performance of the proposed system was evaluated on North Indian Classical music dataset (MIREX-2008) and Carnatic dataset. The performance is comparable to other magnitude spectrum based approaches. Important feature of the proposed algorithm is that it neither requires any substantial prior knowledge of the structure of musical pitch nor any classification framework.