Q-MMT: Qinqiang Opera Generation Based on Multi-track Music Transformer

Qinqiang Opera is a treasure of Chinese intangible cultural inheritance. To inject new vitality into the creation and development of Qinqiang opera, this paper proposes an automatic Qinqiang opera generation method based on deep learning techniques. Specifically, we first collected a new Qinqiang Opera Dataset (QOD) containing 172 Qinqiang opera segments, which facilitates the subsequent research on automatic Qinqiang opera generation. Then, we propose a novel Qinqiang opera generation model named Q-MMT, which is a decoder-only transformer with multiple feed-forward heads and an improved attention mask. Experimental results show that Q-MMT surpasses previous transformer-based multi-track music generation methods in both objective and subjective evaluations, and is able to generate multi-track music with the unique characteristics of Qinqiang opera.