YuE2 makes a readable score part of full-song generation
In plain words
Some music generators make a score you can read but stop before producing a recording. Others make recordings without showing the musical decisions behind them. YuE2 is one system that writes a score, adds coded musical detail, and then makes a full song.
In a test of 192 song requests, expert listeners compared planned songs with songs made by the same system without a score. Among judged overall-quality responses, more favored the planned songs than the unplanned ones. That finding applies to this test and its listeners.
The research
A music generator can make a finished recording while hiding how it chose melody and harmony. Could one system expose those choices in a readable score and still turn them into a complete song? The paper tests whether generating that score first changes expert judgments of the result.
YuE2 generates a readable score before musical tokens and audio, letting one model expose composition and render a full song.
The paper asks whether a music generator can make its compositional choices readable while still producing a finished recording. YuE2 does this by generating a score first, then adding semantic music tokens and acoustic detail to render the song.
In a same-checkpoint comparison across 192 prompts, experts more often favored songs made with melody-and-chord planning than songs made without it: 49.3% versus 34.6% of 211 judged overall-quality responses, with the rest ties.
A separate comparison also favored YuE2 over a separate-model system when both used planning. These results support the score-first approach in the tested comparisons.
These preference results describe the tested prompts, songs, expert listeners, and comparison procedures. They do not show that every listener or music request will favor planned generation.
Analysis
Deep
- Authors
- Ruibin Yuan, Jiahao Pan, Junyan Jiang, Zhiyue Wu, Ziya Zhou
- License
- arXiv non-exclusive