YuE2, MERT2, and SheetSage2 models achieve new benchmarks in AI music generation and analysis

YuE2 Sets New Music Generation Benchmark

The YuE2 model (best-of-8) achieved a score of 6.9632 on SongBench, a benchmark for music generation. This result represents the highest observed mean among 15 evaluated settings on WildSongBench, surpassing Suno v5, which scored 6.8721 in the same comparison. YuE2's performance indicates progress in generating musical compositions.

MERT2 Achieves State-of-the-Art in Audio Understanding

MERT2-30s and MERT2-FS (full-song) models have established new state-of-the-art results across 14 of 15 MARBLE metrics. These metrics cover various aspects of audio understanding, including tagging, key, genre, and emotion recognition. MERT2-30s achieved 91.72 in genre accuracy on GTZAN, and MERT2-FS scored 67.05 in key refined accuracy on GiantSteps, demonstrating advancements in how AI interprets musical content.

SheetSage2 Leads in Music Transcription

SheetSage2 has achieved state-of-the-art performance on 10 of 13 benchmark metrics for music transcription. This single model handles six transcription tasks: beat, downbeat, key, chord, structure, and melody. Notable scores include 82.51 for vocal melody pitch-class note F1 on RWC-Pop and 90.08 for chord recognition (Maj/min) on osu2017, indicating improved accuracy in converting audio to symbolic musical notation.

Impact on AI Music Development

These benchmark achievements across YuE2, MERT2, and SheetSage2 signify progress in AI's ability to generate, understand, and transcribe music. The results contribute to the development of more sophisticated AI tools for music creation, analysis, and research, potentially influencing future applications in the music technology sector.