Color extraction from lyrics

Abstract

We establish a procedure for color extraction from lyrics based on the distributed semantic representation obtained by word2vec. To extract colors from lyrics, we exploit the fact that the cosine similarities between words and color names tend to reflect the actual colors of the things the words indicate. We select important words that characterize the lyrics according to the tf-idf statistic and estimate the colors of the important words using the cosine similarities between the words and color names. For evaluation of the color extraction procedure, we implement a software system that automatically generate images based on the colors extracted from given English lyrics. We conduct a subjective evaluation of the color extraction procedure using the software system and the lyrics of three Beatles songs.