Conditional LSTM-GAN for Melody Generation from Lyrics

Melody generation from lyrics has been a challenging research issue in the field of artificial intelligence and music, which enables to learn discover latent relationship between interesting accompanying melody. Unfortunately, limited availability paired lyrics-melody dataset with alignment information hindered progress. To address this problem, we create large consisting 12,197 MIDI songs each melody through leveraging different music sources where syllables attributes is extracted. Most importantly, propose novel deep generative model, conditional Long Short-Term Memory - Generative Adversarial Network (LSTM-GAN) for lyrics, contains LSTM generator discriminator both conditioned on lyrics. In particular, lyrics-conditioned given notes predicted are generated simultaneously. Experimental results have proved effectiveness our proposed lyrics-to-melody plausible tuneful sequences can be inferred