An improved Kansei-based music retrieval system with a new distance in a Kansei space

The previously proposed music retrieval system is improved. In order to describe individual Kansei traits and in order to evaluate tunes, scores for 40 pairs of Kansei words (adjectives) were used. Fifteen subjects participated in an experiment in which they rated a seven-point scale for each of 12 tunes. The neural network built in the system learned associating each person´s Kansei trait with a standard parson´s one. After learning the system could retrieve tunes out of the built-in database, requested by an user in term of scores for 40 word pairs. We show that the system can be improved by selecting a qualified standard evaluator and introducing a new distance for retrieving tunes that well match to users´ request.