YuE2 · Frontier Music with Symbolic Planning

AI-generated music quality now matches or exceeds human-created benchmark scores.

YuE2, a new music generation model, has achieved a SongBench score of 6.9632, surpassing Suno v5's score of 6.8721 on the WildSongBench benchmark. Developed by Tokenwave.AI, YuE2 unifies symbolic and audio music generation by first creating an editable musical score through symbolic planning, then producing vocals and accompaniment. The model, with approximately 3.59 billion parameters, also excels in music representation with its MERT2 model and audio-to-score transcription with SheetSage2. MERT2 achieves state-of-the-art results on 14 out of 15 MARBLE metrics for music understanding tasks. SheetSage2, a single model, is state-of-the-art on 10 out of 13 metrics for six different transcription tasks, including melody and chord recognition. Both models are trained on licensed synthetic data and CC0 music, with a focus on ethical data use. YuE2 allows for creation, covering, and editing of songs.

YuE2's performance indicates a significant leap in AI music generation quality, positioning it as a new frontier model that directly challenges established players like Suno. The integration of symbolic planning allows for editable scores, offering creators a level of control previously unattainable in AI music synthesis. This development suggests a future where AI tools are not just for generating raw audio but for collaborative composition, with editable symbolic representations at the core.

The underlying advancements in music representation (MERT2) and audio-to-score transcription (SheetSage2) are crucial enablers of YuE2's capabilities. MERT2's state-of-the-art performance across numerous music understanding metrics, and SheetSage2's accuracy in deconstructing songs into editable formats, highlight a maturing AI infrastructure for music. This combination points to an accelerating trend where AI models can both understand and generate complex musical structures with increasing fidelity and controllability.