We release lyrics and audio files used in the qualitative evaluation[1].

For the evaluation sets, we randomly selected four melodies from the RWC music database[2].
For each melody, we prepared four lyrics generated by the Heuristic, Lyrics-only, Fine-tuned, and Pseudo-melody models.
Moreover, to obtain an upper bound for this evaluation, we used the lyrics created by amateur human writers: we asked four native Japanese speakers to write lyrics on the evaluation melody.
One writer was a junior high school teacher of music who had experience in music composition and writing lyrics.
Three writers were graduate students with different levels of musical expertise.
Two of the three writers had experience with music composition, but none of them had experience with writing lyrics.

• [1] Kento Watanabe, Yuichiroh Matsubayashi, Satoru Fukayama, Masataka Goto, Kentaro Inui and Tomoyasu Nakano. A Melody-conditioned Lyrics Language Model. In Proceedings of the 16th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT 2018) In Proceedings of the 3rd of International Society for Music Information Retrieval (ISMIR 2002)
• [2] Masataka Goto, Hiroki Hashiguchi and Takuichi Nishimura and Ryuichi Oka. 2002. RWC music database: Popular, classical and jazz music databases. In Proceedings of the 3rd of International Society for Music Information Retrieval (ISMIR 2002)