Journal for Research | Volume 03 | Issue 05 | July 2017 ISSN: 2395-7549
Finding Similarity between Ragas using KMeans Clustering Bhavana Tiple Research Scholar Department of Computer Science & Engineering G. H. Raisoni COE, Nagpur India
Prof. Achamma Thomas Head of Dept. Department of Computer Science & Engineering G. H. Raisoni COE, Nagpur India
Abstract Data mining is a concept of extracting or mining knowledge from large amount of data. Clustering is a data mining technique in which it is used to grouping the similar data items. K-means is one of the simplest unsupervised learning algorithms that solve the well-known clustering problem. We use K-means to find out the similarities between Ragas, which is the fundamental unit of Indian Classical Music. We estimate the similarity of the example music file and the samples in the queried database by calculating the Euclidean distance and applying K-Means between the features extracted from the music files. Keywords: Indian Classical Music, K-means, MFCC, Raga, Similarity Measure, ZCR _______________________________________________________________________________________________________ I.
INTRODUCTION
Audio Data Mining is an emerging field of research. With Music files being created and generated at a very high rate and also existing huge music data, makes it essential that efficient algorithms be used to store, retrieve, and recommend music files. Finding similarities between the music files is one such task. Finding similarities between various data items, is one of the most prominent task in Data Mining. Work in the direction of finding similarities between various types of music is worked upon by various researchers. A raga is characterized by several attributes, like its Vaadi-Samvaadi, Aaroha-Avaroha and Pakad, besides the sequence of notes which denotes it. No two ragas are same even if they are performed by same artist. There are many minor variations in raga which make them similar or dissimilar. Therefor finding the similarity of raga is an important area in research of raga studies and raga classification, thus we chose raga datasets to find similarity measures in classical Indian music. Raga is the basic unit of Indian Classical Music. Identifying the Raga on which a song is based is a difficult task. Only people who are rigorously trained in Indian Classical Music can accurately identify the Raga With over more than 500 Ragas, it becomes very difficult to identify the raga on which various songs are based on. The specific notes within a raga can be reordered and improvised by the musician. Each raga has an emotional significance and symbolic associations such as with season, time and mood. The raga is considered a means in Indian musical tradition to evoke certain feelings in an audience. The applications of this system involves clustering the ragas which may be further used to identify the raga on which a particular song is based. It can also be used to recommend the ragas which are similar to each other and might interest the user. II. RELATED WORK Lot of work has been done in finding similarity in western music. Some work in raga identification in Indian classical music and Carnatic music is also performed in this section we briefly explain the work done in field of similarity search and recommendation systems created based on this similarity techniques. Min Woo Park et al [1] worked on overall similarity between two different songs. They stated that overall region of a song is represented as 1-dimensional signal which can be obtained by run-length representation of 2-dimensional note information (pitch, duration)). They combined two distance concepts such as the Euclidean distance and the Hamming distance to develop new distance metric to find similarity. Tao Li et al [2] investigates the use of acoustic based features for music information retrieval. Two specific problems are studied: similarity search (searching for music sound files similar to a given music sound file) and emotion detection (detection of emotion in music sounds). For similarity search, the distance between two sound files is defined to be the Euclidean distance of their normalized representations. Based on the distance measure the closest sound files to an input sound file is obtained. Emotion detection is defined as a multiclass classification problem, and is resolved with the use of Support Vector Machines trained on the extracted features. Dominik Schnitzer[3] worked on fiter-and-refine method to speed up nearest neighbour searches for multivariate Gaussians. The combination of features and similarity estimation is widely used to compute music similarity. This paper proposes a method for fast nearest neighbor retrieval in large databases. Jamil George et al [4] proposed a system for automatic interpolative playlist generation based upon a sub symbolic music similarity measure, allowing users to select a set of key tracks, Adam Berenzweig[5] worked on both acoustic and subjective approaches for calculating similarity between artists, comparing their performance on a common database of 400 popular artists. They evaluated acoustic techniques based on Mel-frequency cepstral coefficients and an intermediate anchor space of genre classification. Sankalp Gulati[6] performed a comparative evaluation of methodologies for computing similarity between short-time melodic fragments of audio recordings of Indian art music. Melody representation, pitch quantization levels,
All rights reserved by www.journal4research.org
1
Finding Similarity between Ragas using K-Means Clustering (J4R / Volume 03 / Issue 05 / 001)
normalization techniques and distance measures. Melodic fragment similarity is particularly sensitive to distance measures and normalization techniques III. PROPOSED SYSTEM
Fig. 1: System Architecture of the system
The overall architecture of the system is given in figure1.It has two main processes, Front end User Interface and backend processing. The modules in User interface are Wav/mp3 file loader: It browse the audio raga input. It inputs the audio clip in mp3 or wav format. K-selector: It gives the choice to select the vaues of k ranging from 1 to 5.. K is the number of clusters to be defined. Graph plotter: Each raga is plotted on the graph with 2 factors as x,y axes. Coordinates: Each cluster has a different colour. Table viewer: The cluster information of each raga is displayed in a table format. The user can search any particular raga through the search option The next important task in the similarity search is backend processing. It has following modules in back end Back-end handling 1) Mp3 to wav converter: It is mp3 to wav file conversion module. Mp3 is a lossy format to store an audio file. Some important information may lost in this format. Therefore the audio clips are converted to Wav formt which is a lossless storage of sound clps. 2) File Processing: In this module stereo channels are converted into mono channel. The processing is done in mono channel. 3) Feature extraction: Feature extraction is one of the important process which decides the performance of system. Choice of proper features is very important for generation of feature vector and data values. In this system for 12 MFCC and ZCR features are extracted using R libraries. 4) Storing of extracted features in database: The extracted features are stored in MySQL database. For every new audio input its features are appended to the existing database. 5) Factoring of all the features: The feature set of each raga is a large vector so we do factoring to extract two factors to plot each raga on the graph. 6) K-means clustering algorithm implementation: The data given by factoring are clustered by the k-means method, which aims to partition the points into k groups such that the sum of squares from points to the assigned cluster centres is minimized. Clustering is a data mining technique in which it is used to grouping the similar data items. K-means is one of the simplest unsupervised learning algorithms that solve the well-known clustering problem. We use K-means to find out the similarities between Ragas, which is the fundamental unit of Indian Classical Music. We estimate the similarity of the example music file and the samples in the queried database by calculating the Euclidean distance and applying K-Means between the features extracted from the music files. 7) Cluster Graph and table generation: The graphs are plot according to the similarity of ragas. 8) Dataset: there is no standard database for Indian classical music processing. We have collected database from you tubes and generated from singers. All rights reserved by www.journal4research.org
2
Finding Similarity between Ragas using K-Means Clustering (J4R / Volume 03 / Issue 05 / 001)
IV. CONCLUSION AND FUTURE SCOPE This system has the capability to measure the similarities between note sequences of two or more ragas. The features of the ragas are used to group the ragas and measure the similarities across different ragas. The ultimate goal of the system is to group the similar ragas into clusters using k-means algorithm. Future work lies in proposing a well-defined and robust machine learning technique for measuring the similarities across different ragas which gives maximum accuracy. More robust dataset can be used. ACKNOWLEDGEMENT I would like to very earnestly thank my research Guide Dr M S Patwardhan for her consistent support, invaluable suggestions and
throughout contribution towards the betterment of this work. I sincerely appreciate her continuous guidance and would continue to work under her for producing quality research that would ultimately be socially acceptable to everyone. REFERENCES [1]
Min Woo Park and Eui Chul Lee,” Similarity Measurement Method between Two Songs by Using the Conditional Euclidean Distance”, E-ISSN: 2224-3402 Issue 12, Volume 10, December 2013. [2] Tao Li, Mitsunori Ogihara, “Content-based music similarity search and emotiondetection”,2004 IEEE. [3] Dominik Schnitzer, Arthur Flexer, Gerhard Widmer,”A fast audio similarity retrieval method for millions of music tracks.”,21 December 2010. [4] Jamil George Bashi, “Music Similarity Measures for Interpolative Playlist Generation “, 2013. [5] Adam Berenzweig, Beth Logan, Daniel P.W. Ellis, and Brian Whitman, “A Large Scale Evaluation of Acoustic and Subjective Music Similarity Measures”, 2013. [6] Sankalp Gulati, Joan Serrà, and Xavier Serra,”An evaluation of methodologies for melodic similarity in audio recordings of indian art music”, Music Technology Group, Universitat Pompeu Fabra, Barcelona, Spain Artificial Intelligence Research Institute (IIIA-CSIC), Bellaterra, Barcelona, Spain, IEEE 2015. [7] Maria Panteli, Emmanouil Benetos, Simon Dixon, “Learning a feature space for similarity in world music”, 17th ISMIR Conference, New York City, USA, August 7-11, 2016. [8] Maria Panteli, Simon Dixon,”On the evaluation of rhythmic and melodic descriptors for music similarity”, 17th ISMIR Conference, New York City, USA, August 7-11, 2016. [9] Shu Tamura Shinichi Ito Momoyo Ito, Minoru Fukumi, “Method to Evaluate Similarity of Music by Music Features”, IECON2015-Yokohama November 912, 2015. [10] Dmitry Bogdanov, Joan Serrà, Nicolas Wack, Perfecto Herrera, and Xavier Serra, “Unifying Low-Level and High-Level Music Similarity Measures”, IEEE August 2011.
All rights reserved by www.journal4research.org
3