YuE realizes a technological breakthrough in high-quality music generation
YuE ensures the quality advantage of the generated music through a number of technological innovations. Its semantically enhanced audio disambiguator is able to accurately capture the subtle features of music signals, and its three-phase training scheme effectively solves the problem of modeling long sequences, so that the generated songs remain highly coherent in the time dimension.
In terms of diversity, the model supports the switching of various music styles such as pop and metal, and can adjust the vocal type according to the demand. Tests show that YuE-generated songs significantly outperform traditional methods in the following metrics: musical structural integrity is improved by 471 TP3T, melodic naturalness is improved by 321 TP3T, and style fit is enhanced by 281 TP3T.
These advantages stem from two key points in the design of the model architecture: first, the dual-partitioning technique handles musical and linguistic features separately, avoiding information confusion; second, the lyrics chain thinking simulates the process of human creation, understanding the emotion of the lyrics before generating the corresponding melody, and this hierarchical processing ensures the accuracy of the artistic expression.
This answer comes from the articleYuE: Transforms lyrics into a base model of a complete song, supporting a wide range of musical stylesThe
