LoveLive! Published an AI paper: generating models to write music scores automatically

lately , Preprint paper platform arXiv A paper on attracted people's attention , The author is from Game Developer KLab And Kyushu University . They proposed a model for automatically writing scores for idol songs , what's more , The author said that this method has been used for a long time .

Reducing the cost of music score generation is an important challenge for online music game developers , Because it is the bottleneck of daily operation .KLab The proposed method realizes that only audio is needed , You can directly generate music scores .

In the course of research , The developers first proposed Dance Dance Convolution (DDC) , Has produced a high level of human , Music score of higher difficulty game mode , But the effect of low difficulty is not good . Then the researchers improved the data set and multi-scale conv-stack framework , Successfully captured the time dependence between the quarter notes and the position of the eighth note and the hint beat in the score , They are a good time to place keys in the audio game .

DDC It consists of two sub models ：onset（ The timing of generating notes ） and sym（ Determine the note type , Such as tapping or sliding ）

Currently in use AI The model has achieved good results in all difficult scores , The researchers also look forward to the possibility of extending the technology to other fields .

Thesis link ：

https://arxiv.org/abs/2202.12823

KLab Apply depth generation model to synthesize music score , And improve the production process of music score , Cut business costs by half . This study illustrates how to use a new multi-scale model dedicated to rhythmic movements GenéLive!, Overcome challenges with beats, etc , And use KLab The production data set and open data set were evaluated .

GenéLive! The model architecture of .