Training-based semantic descriptors modeling for violin quality sound characterization