This paper describes our development of a deep learning based time-span tree analyzer of the Generative Theory of Tonal Music (GTTM). Construction of a time-span tree analyzer has been attempted several times, but most previous analyzers performed very poorly, while those that performed relatively well required parameters to be manually adjusted. We previously proposed stepwise reduction for a time-span tree, which reduces the branches of the tree one by one, and confirmed that it can be learned by using the Transformer model. However, stepwise reduction could not obtain a time-span tree because it does not know to which notes the reduced notes were absorbed. Therefore, we improved the encoding for learning stepwise reduction and specified which notes are absorbed by which notes. We also propose a time-span tree acquisition algorithm that iterates stepwise reduction by representing the time-span tree as a matrix. As a result of experiments with 30 pieces, correct time-span trees were obtained for 29 pieces.