YuE2是一套集音乐生成与编辑于一体的前沿系统，在SongBench基准测试中以6.9632的得分超越Suno v5的6.87211。该系统的核心由两个专门模型组成：MERT2音乐编码模型在MARBLE基准的15项指标中达成14项最优性能，SheetSage2转录模型则在节拍、降拍、音阶、和弦、结构、旋律等6项转录任务的13项指标中实现10项最优性能1。
YuE2通过符号化规划支持音乐创作与编辑1。系统的训练基于大规模数据集，其中MERT2使用70万小时训练数据，SheetSage2使用2.84万小时，YuE2本体使用34.6万小时1。这些训练数据主要来自CC0许可音乐和合成数据，后者由Tokenwave.AI提供1。
YuE2, an advanced music generation and editing system, has set new performance records across multiple music AI benchmarks.1 The system achieved a score of 6.9632 on SongBench, surpassing Suno v5's score of 6.8721.1 This represents a significant advancement in the field of AI-driven music creation, powered by symbolic planning capabilities that enable structured musical composition and editing workflows.
The YuE2 platform integrates two core models that drive its performance.1 MERT2, the music encoding model, attained state-of-the-art results on 14 out of 15 metrics within the MARBLE benchmark, while SheetSage2, the transcription model, achieved state-of-the-art performance on 10 out of 13 metrics across six transcription tasks—including beat tracking, downbeat detection, scale recognition, chord identification, structure analysis, and melody extraction.1 The training infrastructure supporting these components involved substantial computational resources, with MERT2 trained on 700,000 hours of audio data, SheetSage2 on 28,400 hours, and the full YuE2 system on 346,000 hours.1 The training datasets primarily consisted of CC0-licensed music and synthetic data generated by Tokenwave.AI.1
评论
还没有评论，欢迎留下第一条。