Identification of Individual Melodies Using Artificial Neural Network Classifier

Abstract

New trends in music distribution and storage have led to tremendous interest in the processing of music signals. From surfing music collections, to discovery new performers, and preventing the records from privacy computers have played a very vital role, thus processing of music is need of the hour. Every part of music can be classified on the basis of genre, artist or one of the many other parameters. The music samples have wide variety of information and details which are very important for different kind of applications. There are various properties of audio signals which are defined such as fundamental periodicity, known as pitch, and amount of overlapping of music instruments in a sample, known as polyphony, and types of characteristics, known as timbre. Some of the properties are just defined in lay-man language but do not have exact mathematical definition, timbre is not defined mathematically. In this report feature extracted and used for processing is chroma feature. Artificial neural networks(ANN) have been applied to many research fields like speech recognition, classification of cancers and gene prediction. In this dissertation ANN is used on music melody samples. ANN is made to learn the data and its formation and then it is tested. Two techniques of ANN, namely Back Propagation Algorithm and Radial Basis Function, are used for this purpose. The chromagrammed data is used as as learning and testing data. Principal Component Analysis (PCA) and Continuous Wavelet Transform (CWT) are both used typically used for large set of data. These two techniques are used as pre-processing techniques for the data set. A comprehensive analysis was carried out in between the learning error rate performance and testing performance of all the pre-processed and non-processed data sets. The comparative results demonstrate the effectiveness of the proposed method.New trends in music distribution and storage have led to tremendous interest in the processing of music signals. From surfing music collections, to discovery new performers, and preventing the records from privacy computers have played a very vital role, thus processing of music is need of the hour. Every part of music can be classified on the basis of genre, artist or one of the many other parameters. The music samples have wide variety of information and details which are very important for different kind of applications. There are various properties of audio signals which are defined such as fundamental periodicity, known as pitch, and amount of overlapping of music instruments in a sample, known as polyphony, and types of characteristics, known as timbre. Some of the properties are just defined in lay-man language but do not have exact mathematical definition, timbre is not defined mathematically. In this report feature extracted and used for processing is chroma feature. Artificial neural networks(ANN) have been applied to many research fields like speech recognition, classification of cancers and gene prediction. In this dissertation ANN is used on music melody samples. ANN is made to learn the data and its formation and then it is tested. Two techniques of ANN, namely Back Propagation Algorithm and Radial Basis Function, are used for this purpose. The chromagrammed data is used as as learning and testing data. Principal Component Analysis (PCA) and Continuous Wavelet Transform (CWT) are both used typically used for large set of data. These two techniques are used as pre-processing techniques for the data set. A comprehensive analysis was carried out in between the learning error rate performance and testing performance of all the pre-processed and non-processed data sets. The comparative results demonstrate the effectiveness of the proposed method.