S`Q+X Q7 i?2 2nd *QKTJmbB+ qQ`Fb?QT UAbiM#mH- hm`F2v- CmHv Rk@Rj- kyRkV CLASSIFICATION OF INDIAN CLASSICAL VOCAL STYLES FROM MELODIC CONTOURS Amruta Vidwans, Kaustuv Kanti Ganguli and Preeti Rao Department of Electrical Engineering Indian Institute of Technology Bombay, Mumbai-400076, India {amrutav, kaustuvkanti, prao}@ee.iitb.ac.in ABSTRACT characteristic phrases rather than just the scale intervals [4]. Computational approaches have not been applied to A prominent categorization of Indian classical music is style discrimination however. Liu et al. [5] attempted to the Hindustani and Carnatic traditions, the two styles classify audio signals according to their cultural styles as having evolved under distinctly different historical and Western or non-Western by the use of characteristics like cultural influences. Both styles are grounded in the me- timbre, rhythm and musicology-based features. More re- lodic and rhythmic framework of raga and tala. The cently, Salamon et al. [6] classified Western genres using styles differ along dimensions such as instrumentation, melodic features computed from pitch contours extracted aesthetics and voice production. In particular, Carnatic from polyphonic audio. music is perceived as being more ornamented. The hy- Hindustani and Carnatic music differ in the nature of pothesis that style distinctions are embedded in the me- the accompanying instrumentation and can potentially be lodic contour is validated via subjective classification distinguished by acoustic features relating to timbre. tests. Melodic features representing the distinctive char- However, it may be noted that the two styles can also be acteristics are extracted from the audio. Previous work reliably distinguished by listeners of the vocal music ex- based on the extent of stable pitch regions is supported by tracted from the alap section (i.e. the improvised compo- measurements of musicians‘ annotations of staEle notes. nent) of a concert where the accompanying instrument is Further, a new feature is introduced that captures the restricted to the common drone (tanpura). A common presence of specific pitch modulations characteristic of perception among listeners is that the Hindustani alap ornamentation in Indian classical music. The combined unfolds —slowly“ relative to the corresponding Carnatic features show high classification accuracy on a database alap which has complex pitch movements (gamakas) [7]. of vocal music of prominent artistes. The misclassifica- These observations imply that the melodic contour of the tions are seen to match actual listener confusions. alap contains sufficient information about style differ- ences. In this work we consider the automatic identifica- 1. INTRODUCTION tion of the style (Hindustani or Carnatic) from the melod- ic contour. Since transcriptions in the form of symbolic Indian Classical Music styles span a wide range, a promi- notation are not easy to come by (apart from the absence nent categorization within which is Hindustani and Car- of standard notation to represent pitch movements), we natic. The distinction is geographical with the two styles investigate style recognition from the available recorded having evolved under distinctly different historical and audio of vocal performances. Such work can be useful in cultural influences. Carnatic music is predominantly per- providing musicological insights as well as in developing formed and studied in the southern states of India while tools for music retrieval. Hindustani music is more widely spread in the country. The repertoire of commonly performed ragas differs Both styles are grounded in the melodic and rhythmic in the two styles. However, in order to minimize any ra- framework of raga and tala. While the repertoire of ga-specific influence on the discriminatory characteristics commonly performed ragas is different in the two styles, of the melodic contour in the present study, we choose they share the basic scale structure, the use of raga- music belonging to corresponding ragas in the two vocal specific phrase motifs and ornamentation. In both styles styles. We examine the assumption that the style distinc- due importance is accorded to both compositions and im- tions are represented in the melodic contour via listening provisation although the relative weighting tends to dif- tests. Next discriminatory features that can be computed fer. The styles differ along dimensions such as structure from the detected pitch contour are presented and evalu- of a performance, aesthetics, voice production and the use ated for automatic style identification. of decorative elements. Additionally, Hindustani and Carnatic styles differ in the musical instruments used. There has been some past work on the computational 2. MELODIC FEATURE EXTRACTION analysis of Indian classical music related to automatic In order to characterize the melody, it is necessary to first recognition of raga [1, 2, 3]. These approaches have been extract it from the polyphonic audio signal. The accom- based on the distinctness of scale intervals, precise into- panying instrument in the alap section of the concert is nation and phraseology. With a raga being far more con- restricted to the tanpura (drone). Melody detection in- strained than the Western scale, its grammar is defined by volves identifying the vocal segments and tracking the pitch of the vocalist. Indian classical singing is a pitch- Copyright: © 2012 Amruta Vidwans et al. This is an open-access article dis- tributed under the terms of the Creative Commons Attribution Licen- RjN se 3.0 Unported, which permits unrestricted use, distribution, and repro- duction in any medium, provided the original author and source are cre- dited. S`Q+X Q7 i?2 2nd *QKTJmbB+ qQ`Fb?QT UAbiM#mH- hm`F2v- CmHv Rk@Rj- kyRkV continuous tradition characterized by complex melodic 2.2 Musically motivated features movements. These ornamentations (gamak) are catego- rized based on shape into a variety of glides and oscillato- Carnatic vocal renditions are typically replete with orna- ry movements. The oscillatory movements include sever- mentation as opposed to the relatively slowly varying al that are slower in rate and larger in amplitude than the pitches of the Hindustani vocalist. The difference is par- Western vibrato. In this section, we present the imple- ticularly prominent in the alap section which the artiste mentation of vocal pitch detection in such a scenario fol- uses for raga elaboration and where the svar appear in lowed by a discussion of melodic features that character- their raga-specific intonation whether steady or orna- ize the style differences. mented with touch notes (kan) or oscillations (gamak). We explore the possibility of a musicologically motivated feature for the above difference. Hindustani musicians 2.1 Vocal pitch detection refer to held notes as —standing“ notes or khada svar. A We employ a predominant-f0 extraction algorithm de- manual annotation of 20 minutes of audio comprising of signed for robustness in the presence of pitched accom- 30 alap sections across different ragas rendered by prom- paniment [8]. This method is based on the detection of inent Hindustani vocalists was carried out by 2 trained spectral harmonics, helping to identify multiple pitch musicians. The musicians labeled the onset and offset of candidates in each 10 ms interval of the audio. Next pitch each instance of khada svar that was perceived on listen- saliency and continuity constraints are applied to estimate ing to the audio. The duration and standard deviation of the predominant melodic pitch. Although the drone is au- each instance was measured. Figure 3 shows scatter plots dibly prominent due mainly to its partials spreading over of the 241 instances of khada svar identified by the musi- the frequency range up to 10 kHz, the strengths of its cians. We observe that the location of the highest density harmonics are low relative to the voice harmonics. Thus is duration=700 ms and standard deviation=10 cents. the singing voice dominates spectrally, and the melody Thus these may be considered as nominal values for a can be extracted from the detected pitch of the predomi- khada svar as obtained by this experimental investiga- nant source in the 0-4 kHz range. tion. In the next section, we propose a method to seg- State-of-the-art pitch detection methods achieve no ment the pitch contour into steady and ornamented re- more than 80% accuracy on polyphonic audio. An im- gions depending on the detected local temporal variation portant factor limiting the accuracy is the fixed choice of [12]. spectrum analysis parameters, which ideally should be matched to the characteristics of the audio such as the 2.3 Stable note segmentation pitch range of the singer and the rate of variation of pitch. Steady, or relatively flat, pitch regions are expected to In the regions of rapid pitch modulation, characteristic of correspond to the khada svars of the underlying raga. Indian classical singing, shorter analysis windows serve Based on the observations of the previous section, a sta- better to estimate the vocal harmonic frequencies and ble note region is defined as a continuous segment of a amplitudes. Hence for better pitch detection accuracy, it specified minimum duration (—1“ ms) within which the is necessary to adapt the window length to the signal pitch values exhibit a standard deviation less than a speci- characteristics. This is achieved automatically by the fied threshold (—-“ cents) from the computed mean of the maximization of a signal sparsity measure computed at segment. Figure 1 depicts the detected steady note segments as each analysis instance (every 10 ms) of local pitch detec- dark lines superposed on the continuous pitch contours tion [9]. Finally, it is necessary to identify the vocal re- using the nominal parameters N=400 ms and J=20 cents. gions in the overall tracked pitch. This is achieved by us- The gamakas, or complex pitch movements, are left un- ing the peculiar characteristics of Hindustani music where touched. We observe that the long held notes coincide the vocal segments are easily discriminated from the in- with the svar locations of the raga.