SongGeneration 2 beats Suno on lyric accuracy

Open-source music AI just crossed the commercial quality threshold — and the gap to Suno and Udio is now measurable and real.

• –PER benchmark puts SongGeneration v2 ahead of every commercial competitor: 8.55% vs. Mureka v8's 9.96% and Suno v5's 12.4% — lyric hallucination has been the Achilles heel of open-source music AI, and DPO training on 200K preference pairs finally kills it
• –Hybrid LLM-Diffusion architecture uses dual parallel token streams (mixed + separated vocal/accompaniment) for structural coherence, then a diffusion renderer for high-fidelity audio — smart two-stage design borrowed from image gen playbook
• –Three model variants span 10–28GB VRAM, making it accessible across consumer and datacenter GPUs; the large model runs 0.82x real-time on an H20
• –Multilingual from day one (Chinese, English, Spanish, Japanese) with structured lyric tags for verse/chorus/bridge — gives producers actual compositional control, not just vibes prompting
• –CC BY 4.0 license and local deployment mean zero per-generation cost; community already shipped UI wrappers and WaveSpeed added hosted API at $0.05/gen within days of launch