I-Vectors for Timbre-Based Music Similarity and Music Artist Classification

In this paper, we present a novel approach to extract songlevel descriptors built from frame-level timbral features such as Mel-frequency cepstral coefficient (MFCC). These descriptors are called identity vectors or i-vectors and are the results of a factor analysis procedure applied on framelevel features. The i-vectors provide a low-dimensional and fixed-length representation for each song and can be used in a supervised and unsupervised manner. First, we use the i-vectors for an unsupervised music similarity estimation, where we calculate the distance between i-vectors in order to predict the genre of songs. Second, for a supervised artist classification task we report the performance measures using multiple classifiers trained on the i-vectors. Standard datasets for each task are used to evaluate our method and the results are compared with the state of the art. By only using timbral information, we already achieved the state of the art performance in music similarity (which uses extra information such as rhythm). In artist classification using timbre descriptors, our method outperformed the state of the art.