StepAudio 3 Music plans songs with ABC notation, ranks #3 behind Suno V5.5
stepfun-ai · hf · 2026-09-16
StepFun introduced StepAudio 3 Music, a long-form music generation model with explicit musical planning and open-domain text control.
- Architecture: a 50-Hz single-codebook tokenizer (65536 entries) plus a flow-matching DiT predicting continuous VAE latents, decoded to 48-kHz audio; design choices validated against single-codebook VQ and semantic/acoustic RVQ baselines
- ABC-CoT: a MoE autoregressive model first generates an ABC-notation arrangement plan, making harmony, rhythm, and melodic structure part of the generation context
- Capabilities: songs and instrumentals up to 5:30, dry-vocal accompaniment, and cover synthesis; progressive curriculum, SFT, and DPO
- Results: best AudioBox Content Enjoyment/Usefulness/Production Quality scores; Quality Elo 1105 on the Artificial Analysis Music Arena Vocals leaderboard — behind only Suno V5.5 and Mureka, ahead of Suno V5 and MiniMax
Audio demos: stepaudiollm.github.io/step-audio-3-music.
Related event: StepFun Releases Five StepAudio 3 Models Covering Full Audio AI Stack(6 posts)→
More from Multimodal
- A 25-year design veteran asks: who still mixes pro software with AI tools? — Opening-Knee-5913 · 2026-11-02
- Google Pay pilots AI fashion Discover tab in India with virtual try-on and photo search — heyshrutimishra · 2026-11-02
- Reddit user vents on MiniMax: plasticky AI look, 20GB+ models eating SSD space — DelboyTrigger · 2026-11-02
- Paper folk-story prompt template produces charming cut-out collage images — azed_ai · 2026-11-02
- SeedVR2 vs MiniMax Open-H3 vs FlashVSR+ Turbo: real 4K upscale timings on RTX 5090 — CeFurkan · 2026-11-02
- Could one person with Kling and CapCut tell the same story? — tetsuoai · 2026-11-02