Raga Classification for Carnatic Music

Abstract

In this work, an effort has been made to identify raga of given piece of Carnatic music. In the proposed method, direct raga classification without the use of note sequence has been performed using pitch as the primary feature. The primitive features that are extracted from the probability density function (pdf) of the pitch contour are used for classification. A feature vector of 36 dimension is obtained by extracting some parameters from the pdf. Since non-sequential features are extracted from the signal, artificial neural network (ANN) is used as a classifier. The database used for validating the system consists of 162 songs from 12 ragas. The average classification accuracy is found to be 89.5 %.

1 Introduction

Computational Musicology is an emerging field that includes the study of techniques for analyzing various kinds of music. It helps to study science behind music and develops a scientific framework for this art. The analysis of music helps to understand the culture and the heritage of the society the music is evolved in. Most of the studies in this area are dedicated to Western Music whereas very fewer studies have tried to explore Indian Classical Music (ICM) [1, 2]. ICM is broadly classified into Hindustani and Carnatic music. Though these two music types have same base and similar framework, they differ in many factors. ICM is characterized by two basic elements—it should follow a specific note sequence known as raga and a specific rhythm known as Taal [3]. Raga is considered to be the backbone of ICM. In a broader sense, raga can be understood as the framework that tells all the protocols that should be followed while rendering the piece of music. Since raga is the crucial element in the melodic framework of ICM, automatic raga identification is one of the important steps in Computational Musicology as far as ICM is considered. Some of the applications of automatic raga identification are: development of music recommendation systems, automatic note transcription, music indexing, on-line teaching and learning of music and so on. It may be considered as the first logical step in the process of creating computational methods for ICM.

Some of the characteristics of raga are swara (notes), arohana-avarohana pattern and characteristic phrases. Technically, a note can be defined as an identifiable fundamental frequency component (pitch) of a singer with a beginning and an ending time, of a duration determined by these [2]. The ratio of the fundamental frequencies of two notes is referred to as an interval [2]. The notes are categorised as Shadja (Sa), Rishaba (Ri), Gandhara (Ga), Madhyama (Ma), Panchama (Pa), Dhaivatha (Da) and Nishadha (Ni). These seven notes are further classified into different types based on the interval system (scale) used. There are three kinds of scales that are generally used in Carnatic and Hindustani music theory: a 12-note scale, a 16-note scale and the scale which contains 22 microtones. Each raga is defined using unique sequence of notes that is used in rendering the music clip. Few ragas share same set of notes, but they differ in the melody because of different arohana-avarohana pattern and characteristic phrases. Arohana-avarohana pattern refers to ascending and descending progression of notes. This tells about the transition between the notes that can occur in a given raga. Characteristic phrase is defined as a condensed version of the characteristic arrangement of notes, peculiar to each raga, which when repeated in a recital enables a listener to identify the raga being played [3]. Other than these important features, raga is also characterised by the relative position, strength and duration of notes and also gamakas. Gamakas are the beautification elements that tell about the dynamics of the pitch contour.

Unlike western music where the frequency of notes is fixed, ICM gives liberty of choosing the frequency of base note to the artists while performing. Western classical music uses equal tempered notes whereas just-intoned notes are used in ICM. In equal tempered scale all notes are equally spaced i.e. all notes have the same frequency ratio with respect to its previous note. But in just-intonation scale the notes are not equally spaced. In ICM, no two performances of the same raga by the same singer would be same. ICM is relatively unexplored because of its complicated grammars, extensive use of gamaka and the liberty given to singer for improvising while performing. All these complexities in analysing ICM, left it as relatively less explored computational musicology. In ICM, the sequential note information plays an important role in construction of melody unlike chords and rhythm which play an important role in western music. Hence, many of the methods proposed in the past [3–5] have used Hidden Markov Model (HMM) for identification of ragas.

In this work, an effort has been made to identify raga without the knowledge of the scale using low-level features. Pitch and its derivatives are used as features for identification of the raga. Various kinds of musical clips from both Janaka (ragas with all 7 notes) and Janya ragas (ragas having 5 or 6 notes) are used for the study.

The paper is organised as follows. A review of related work is given in Sect. 2. Section 3 explains the proposed method. Experimentation and results along with explanation of database are given in Sects. 4 and 5 concludes the work with some further research directions.

2 Literature Review

Since raga gives melodic framework attained through pitch manipulation, many research contributions use pitch and its derivatives as the major features for the task of raga identification. Some of the works [6, 7] have also used timbre and other features along with pitch derivatives. The literature can be broadly classified into two categories. The first, the works that have used note information explicitly for identification of raga and the second, the works that have not used note information. We discuss both kinds of work in this section. One of the main challenges in reviewing the works is lack of standard corpora of ICM. The works that have been carried out are evaluated using different databases collected by the researchers. Hence while reviewing the work the databases considered are also discussed.

In the work [3], HMM is used for the identification of Hindustani ragas. The method proposed in this work uses note sequence as a feature. Two heuristics namely Hillpeak heuristics and Note duration heuristics are proposed for note transcription. Many micro-tonal variations present in the ICM makes the note transcription a challenging task even for a monophonic piece of music. The heuristics proposed in this method tries to overcome these variations. Hillpeak heuristics considers the changes in the sign of the slope of the pitch contour for identification of presence of a note. Note duration heuristics assumes that a note sustains for the period of at least 25 ms and estimates the note for every 25 ms. Another main contribution of this work is characteristic phrase analysis. A n-gram model is proposed for identification of repetitive sequence of notes and based on this raga is identified. The overall raga identification accuracy is claimed to be 77 %. The main drawback of the work is limited data set and accuracy of note transcription. Only two ragas are used for validating the system. The database has many limitations such as the musical notes in arohana-avarohana sequence are sung externally, the piece of music should be in G-sharp scale. The accuracy of note transcription is not appreciably high even after lot of constraints on the dataset. Similar work has been carried out by Arindam et al. [4]. The note transcription of the given musical piece is manually performed and this sequence is used for evaluating HMM. It is observed that HMM is 100 % accurate in identifying ragas if correct note sequence is given as input. However, it is difficult to achieve high accuracy in ICM transcription because of the micro-tonal variations and improvisations. Rajeswari et al. [8] have used tala information to identify the number of notes in the given musical piece, further this is used for identification of ragas. Surendra et al. [9] have used similar method of identifying the raga in Carnatic music using the note sequence information. Since the data set is restricted to single scale identifying the notes is simpler task. The features extracted are the notes present, number of notes present and type of arohana-avarohana sequence. Arohana-avarohana patterns are classified as linear or nonlinear based on the pair of notes present in the sequence. They have used ANN as a classifier. The system is tested using 20 ragas with 3–5 songs in each raga. The paper has claimed 95 % of accuracy in raga identification. Ranjani et al. [10] have proposed a method for identification of tonic and other notes which leads to identification of raga from the Carnatic music. The property of the tonic note with pitch of highest mean and least variation shown in pitch histogram, is used for identification of tonic pitch value. This method estimates the probability density function (pdf) from the pitch contour. Using Semi-Continuous Gaussian Mixture Model (SC-GMM), tonic frequency and raga of the musical piece are identified. The data set used for this experiment consisted of 5 Sampurna ragas which is the major limitation. This paper has claimed 86.3 % of accuracy in raga identification. However, the experiments conducted were limited only to arohana-avarohana sequences. The main drawback of the method that use note information explicitly is that any error in identification of notes, contributes to the error in identification of raga as well.

The works that do not consider note information, mainly concentrate on variations in pitch histogram for raga classification. The pitch histograms give better results for stable regions of smaller input clips compared to longer duration inputs. In [11], spectrally derived tone profiles are used as features. A tone profile is a histogram of note values weighted by duration [11]. The tone profile is obtained by taking the DFT of the segment and summing up the energies of the bins surrounding the note centers. It is further normalised using tonic frequency. The music clips were classified using k-NN algorithm. This paper has claimed 100 % of accuracy in Hindustani raga identification. However this method cannot be generalised because the data set consisted of very few ragas and the music clips were rendered by single artist. Parag et al. in [12] have used Pitch Class Distribution (PCD) and Pitch Class Dyad Distribution (PCDD) as features. PCD is obtained directly from the pitch histograms, that is normalised using the input tonic value and folded into one octave. PCDD is a note bi-gram which is called dyads in musical terms. Dyads are calculated from rough estimation of notes using pitch onsets. This paper has claimed 97 % raga recognition accuracy. The system is tested using the data set consisting of instrumental music clips played by a single artist in 17 ragas. In the work of Koduri et al. [13], the predominant pitch from vocal region of polyphonic music is used for identification of ragas from Carnatic music. The raga classification is performed using pitch histogram weighted by duration and frequency of occurrence as the features. The raga classification is performed using k-NN classifier. The experiment is performed on the data set consisting of 176 tunes from 10 ragas consisting of both vocal and instrumental music. This method is claimed to give 75.6 % accuracy. Pranay et al. [6] have used the concept of vadi swara (the dominant note for the raga) for identification of raga in Hindustani music. From the studies, it is shown that Hindustani note intervals are nearer to equal tempered than just intonation intervals [14]. In [6], ICM notes are associated with Western note intervals for extracting features. This method uses chromagram features along with MFCCs and timbre features for identification of raga. A chromagram is a visual representation of energies in the 12 semitones (or chromas). The most repeating frequency component or semitone is assumed as the vadi. The classification is performed using HMM and GMM. 97 % accuracy is claimed using this method. Even though the results are promising, it cannot be generalised since the system is evaluated using only 4 ragas. The vadi swara for different ragas cannot be unique when a larger data set with more than 20 ragas is considered.

In this work, an effort has been made to overcome some of the limitations that exist in the past methods. The proposed method considers a comprehensive data set consisting of variety of music clips from different kinds of ragas. We try to eliminate the explicit note identification and scale dependency (tonic identification).

3 Proposed Method

In the proposed method pitch and its derivatives are used as features for identification of raga since melody is the result of variations in the pitch. From [7], it is evident that the pitch histogram for different ragas differ. Hence in this study we use some of the parameters obtained from the pdf of pitch contour for identification of raga. The flow diagram of the proposed system is shown in Fig. 1. In training phase, the extracted features and corresponding raga labels are used to construct a ANN model for classification. In testing phase, the feature vectors are fed to trained ANN model for obtaining the raga labels. The accuracy of the system is obtained by comparing them against the actual raga labels.

Fig. 1

Flow diagram of the proposed system

3.1 Feature Extraction

The pitch contour is extracted from the input signal (piece of music) using autocorrelation method [15]. The average pitch values are obtained for the frame of every 50 ms with 50 % overlapping to avoid transition cliques. The pitch contour of arohana-avarohana sequence is shown in Fig. 2b. Different pitch values corresponding to different notes are rendered in the piece of music. A sharp rise and fall in the contour is observed when the transition from a note to another takes place. From the pitch values, pdf is obtained using kernel density estimation method (Gaussian kernel is assumed). Bandwidth range is taken into account from minimum and maximum pitch values. The pdf obtained is shown in Fig. 2c. One can observe that the prominent peaks correspond to the notes present in the raga. Even though Carnatic music is defined using 16-note interval system, only 12 distinct frequency components are present [14]. Each raga is defined using 5, 6 or 7 note combination. Hence from the pdf curve, parameters of the 12 prominent peaks are considered for characterizing the raga. From the pdf, frequency of 12 peaks are obtained and their mean (height of the peaks in pdf) and variances are calculated. Variance of the peak is calculated by considering 20 frequency values on both side of the peak frequency. Since the database consists of clips rendered by different singers in different scale, there is a need to normalise the frequencies of peaks. The frequency that lies in between 100 and 200 Hz having the highest mean is used as a normalising factor. For some clips played by instruments, all the frequency components lie above 200 Hz. In such cases, the frequency having highest mean is used as a normalising factor. All 12 prominent frequencies are divided by the normalising frequency. The peaks and mean values extracted are illustrated in Fig. 2c.

Fig. 2

Feature extraction for raga Ananda Bhairavi. a A music piece rendered in raga Ananda Bhairavi. b Pitch contour of the input music clip. c Peak selection from the pdf of pitch contour

The feature vector obtained using the above procedure is of 36 dimensions (12 normalised frequency values, 12 mean values of the peak frequency and 12 variances). Features 1–12 are the normalised prominent pitch values, features 13–24 are the height of the corresponding peaks and the features 25–36 are the variance of the corresponding peaks. This is illustrated in the Fig. 2c.

3.2 Raga Identification Model

ANN consists of number of simple processing elements, called neurons, that are interconnected to each other [16]. A typical multilayer network, consists of a number of layers namely input layer, output layer. One single neuron makes the simple operation of a weighted sum of the incoming signals and a bias term (or threshold), fed through an activation function resulting in the output value of the neuron [16].

In the proposed system, ANN with feed-forward backdrop learning algorithm is used as a classifier (shown in Fig. 3a). It consists of 1 hidden layer. The input layer consists of 36 neurons which is equals to the size of feature vector. The number of neurons in the hidden layer is taken as 40. This number is chosen after experimenting with various number of neurons ranging from 20 to 50. The accuracy of the system is maximum when the number is 40. The number of neurons in the output layer is fixed as 1. Activation function for hidden and output layers is chosen as tan-sigmoid and pure-line respectively. For each raga different value is expected as an output. For example, if 2 ragas are considered for classification, if input is raga 1, then 1 is expected as outcome and if input is raga 2, then 2 is expected as output value. ANN is trained accordingly with different labels for different ragas. During testing, the raga label obtained from the system (output label) is validated using the actual raga labels (target labels refer Fig. 3b). The difference between system output and actual output is calculated resulting in the error vector. If the error value is strictly less than 0.5, then the input is considered to be correctly classified. This process is illustrated in Fig. 3b. The gray marked values in error vector indicates mis-classification.

Fig. 3

Raga classification model. a Architecture of ANN. b Error vector calculation and validation

4 Result and Analysis

4.1 Database

Monophonic and polyphonic clips from 12 ragas listed in the Table 1 are considered to form a database. Ragas are chosen in such a way that database includes all notes that are present in Carnatic Music. Particularly, Kalyani and Shankarabharanam both ragas are included because Kalyani when perceived in a different tonic becomes Shankarabharanam. There is high chance of mutual mis-classification between these two ragas if the identification is highly tonic dependent. The proposed method is tonic independent and hence the classification of Kalyani and Shankarabharanam is crucial aspect of this work. In the same way, we also include the ragas which share the same set of notes such as Ananda Bhairavi and Reethigowlai (These two ragas share same set of notes), as there is high level of mis-classification between them if set of notes is used for classification. The polyphonic music clips consist of vocals accompanied by Tambura or Tabla. The database is made rich in variety by including arohana-avarohana, krutis, alapanas from different ragas rendered by instruments and different singers including both male and female. We have also included both Janaka and Janya ragas in our database. Janaka ragas (Sampurna ragas) contain all seven notes whereas Janya ragas contain fewer than seven notes.

Table 1 Database: Ragas used and number of clips

4.2 Performance Evaluation

The experiment is performed using different number of neurons in the hidden layer. Considering 40 number of neurons in the hidden layer yielded the best results. The system is validated using three experiments. Initially the raga identification system is developed with 131 music clips from the first 10 ragas of the Table 1, using 70 % of them for training and 30 % for testing. The average raga recognition accuracy is found to be 91.5 %. The accuracy of recognition for each raga in this experiment is given in Table 2. In the second experiment, the system is tested to verify if the features are able to distinguish between ragas that are composed using the same set of notes (Ananda Bhairavi and Reethigowlai). The system is validated using only these two ragas. Using 21 clips for training and 10 clips for testing the accuracy of classification in this experiment is found to be 80 %. This can be justified using the normalisation performed using the pitch value having highest mean and the other parameters such as mean and variances. This either corresponds to tonic or highly repeating note which is called Jeeva Swara in Carnatic music. Jeeva Swara is different for different ragas when they are composed of same notes. If normalisation factor is tonic, then the means and variances of the peaks differ because the importance given for various notes is different in different ragas. Hence classification of raga becomes efficient using the proposed features.

Table 2 Accuracy of the raga classification

In the third experiment, the raga recognition system is validated using 5-fold validation method. Complete database consisting of 162 clips from 12 ragas is divided into 5 sets, each set consisted of songs from all ragas. From these sets, 4 are used for training the network and the remaining one is used for testing. The experiment is repeated for 5 times so that the system is validated against each set. The average result of 5 experiment for each raga is listed in Table 2. The average classification accuracy is 89.51 %.

5 Conclusion and Future Scope

In this work, a system for direct raga classification without use of explicit note sequence is proposed. Set of low-level features obtained from the signal are used as features. A comprehensive data set consisting of variety of music clips is used for validating the proposed method. An effort has been made to come up with a raga classification system that is independent of tonic note information. From the study, it is evident that pdf and primitive features give good accuracy even for ragas composed of same set of notes. The accuracy of the system clearly demonstrates that the features capture more details about raga rather than just the set of notes.

As a future work, the system shall be tested with larger database having hetero-phonic music clips (composed of more than one melody such as vocal accompanied by violin) and also film songs that are composed based on ragas. The proposed features shall be tested with other classifiers as well. Further, the raga information shall be used for developing systems for transcription of notes, music recommendation based on raga information and so on.