MFCC, Spectral and Temporal Feature based Emotion Identification in Songs

Abstract

This work aims for the solution of one of the challenging yet evolving problem of music information retrieval field i.e. identification of emotions in songs. A collection of four datasets of four separate song sample sizes were selected for the purpose of experiments. For each experiment features such as MFCC, spectral and temporal were extracted for each sample of the dataset. A multilayered sigmoidal feed-forward neural network was trained for construction of a model by using error back propagation algorithm. This helped in recognition of four emotion categories (sad, happy, peaceful and angry) from the song samples. The results obtained at the end of these experiments strongly suggest that this trained model was successfully able to identify the emotions in the selected song samples with an average class accuracy of 88.65%.