An automatic singing impression estimation method using factor analysis and multiple regression

This paper describes a method for estimating the impression of a singing voice via acoustic features. While much research has been conducted on singing impression, to date no method for determining appropriate words to represent the impressions created by a person's singing has been developed, primarily due to the lack of a comprehensive evaluation scale. We followed two steps: construction of such an impression scale, and development of models for estimating the impression score of each word. In the scale construction, two experiments were carried out. Firstly, 44 words were selected as relevant words based on subjective evaluation. Secondly, 12 words were selected as an impression scale, and three factors ("powerful", "cautious", and "cheerful") were extracted by factor analysis. To estimate impression scores, multiple regression models were constructed for each impression word with acoustic features. The models were tested by cross validation. The average R2 value for the 12 words of the complete scale was 0.567, and the R2 for the three factors were 0.863 (powerful), 0.381 (cautious), and 0.603 (cheerful).