> Q5 In the case of Markov, etc., I understand that the generation may be not very conformant to the style learnt, unless using a high order Markov but with the risk of recopying entire sequences from the corpus and thus plagiat. But, in the case of a RNN-based architecture [9], what is the rationale?-------------------- A5: As mentioned before, RNN does the similar work as LSTM in our work. But without including a discriminator, it only learns transmission probability between adjacent notes, but does not promise that generated sequences look like real ones.

Come on, guys, that's just not true. You do not need an adversarial loss to get good quality melodies. Look at Sturm's char-RNN on ABC notation, or OpenAI's MuseNet, or a bunch of Project Magenta work, or my own GPT-2 ABC music (MIDI in progress): https://www.gwern.net/GPT-2-music Or for that matter, any generative model trained with a non-adversarial loss (anything using GPT-2 for example).

In fact, generally, everyone avoids GANs for sequence generation because they work so badly compared to regular likelihood training... (Just at a skim, their 'baseline' is pretty suspicious. I'd expect an ablation for the GAN, not comparing their 400-unit LSTM to... a 100-unit LSTM https://www.aclweb.org/anthology/N19-4015.pdf ? Really?)