← Latest briefing

Technology

Music model YuE2 achieves top evaluation score on SongBench

YuE2 recorded the highest mean score across tested configurations on WildSongBench, according to Hacker News.

The short version

  • YuE2 achieved a score of 6.9632 on SongBench under a best-of-eight setting across 192 prompts, exceeding Suno v5's score of 6.8721.[Hacker News]
  • Related releases include MERT2 models and SheetSage2, which set benchmark records across audio transcription and classification tasks.[Hacker News]
  • The models rely primarily on CC0 music and licensed synthetic data from Tokenwave.AI for training.[Hacker News]

Key facts

  • YuE2 reached a mean score of 6.9632 under a best-of-eight evaluation on WildSongBench across 192 prompts.[Hacker News]
  • Suno v5 achieved a score of 6.8721 within the same WildSongBench comparison.[Hacker News]
  • MERT2-30s and MERT2-FS attained top benchmark performance on 14 of 15 MARBLE metrics.[Hacker News]
  • SheetSage2 led across 10 of 13 benchmark metrics for multi-task music transcription.[Hacker News]
  • Model training utilized CC0 music and licensed synthetic datasets primarily provided by Tokenwave.AI.[Hacker News]

What remains uncertain

  • Reported performance metrics stem from a single source submission and have not been independently validated across wider production workloads.[Hacker News]

Sources

Outlet counts describe coverage, not independent confirmation. Reports may share a wire service or original source.