On multi-aspect classification of music data

Abstract

Automated classification of music signal is an active area of research. It can act as the fundamental step for various applications like archival, indexing and retrieval of music data. In this work we have presented an automated system to classify a music signal based on various aspects like its genre, singer, emotion. For Hindusthani classical music, raga being a crucial property we have worked on raga identification also. The system will have two major modules like feature extraction and classification. Current work focuses mostly on the extraction of suitable low level features that can depict the characteristics meaningfully. For classification we have relied on conventional classifiers. In broad sense genre reflects the style. To capture the characteristics of different genres, music signal is first decomposed to extract the component reflecting the desired degree of local characteristics using empirical mode decomposition (EMD). Pitch based features are then computed from the signal at suitable intermediate frequency range. Experimental results and comparison with other works on benchmark dataset indicate the effectiveness of the methodology. Singer identification or singer based classification of song data is very important in the context of music retrieval. Normally a song is accompanied by background music that may cause hindrance. To address this issue we have presented a simple methodology to extract the vocal dominating segments and also to reduce the impact of the instruments. Keeping the physiological aspects of the voice production process and perceptual aspects of human auditory system in mind, features are designed to represent the voice profile of a singer. Spectrogram based vocal-print is proposed to capture the salient timbral characteristics. Mel-frequency cepstral coefficients (MFCC) based features are used as supplement. Emotion being a perceptual and subjective concept, classification based on emotion of music is quite challenging. It is very difficult to design the low level descriptors to represent the emotion. Two methodologies are detailed in this work. In the first approach, a large feature set is considered. The set includes time domain features, spectral features, linear predictive coding and MFCC based features. Different classifiers like, neural network, support vector machine and random forest are tried. In general the performance of such approaches is limited. It is difficult to obtain a consistent feature set that works across the classifier and datasets. To get rid of these issues, deep learning based approach is tried. A conventional neural network built around VGGNet is proposed. It provides substantial improvement of performance. For Indian classical music, raga is the basic melodic framework. Manual identification of raga demands high expertise which is not available easily. Thus an automated system for raga identification is of great importance. In this work, we have studied the basic properties of the ragas in North Indian (Hindusthani) classical music and designed the features to capture the same. Pitch based Swara (note) profile is formed. Occurrence and energy distribution of notes generated from the profile are used as features. Note sequence plays an important role in the raga composition. Proposed note co-occurrence matrix summarizes this aspect. Finally, an experiment has been done for multi-aspect classification. It simply combines the methodologies developed for individual attribute based classification. Previous methodologies have been tried on datasets which are benchmarked for individual aspect. A database has been created to study the performance of multi-aspect classification strategy. Experiment on it shows that proposed methodology performs satisfactorily.