music generation — two paradigms, one legal landscape
A. AR token LM over neural codec
MusicGen · ACE-Step · YuE · VALL-E-Music
text prompt → T5 / CLAP embedding
→ AR transformer over EnCodec tokens
→ EnCodec decoder → audio 32 kHz
streams (future); structured songs with lyrics
drift past 30 s; loops; needs crossfade
B. latent diffusion over mel / latent codec
Stable Audio · AudioLDM 2 · Riffusion
text → CLAP embedding
→ latent diffusion (DDPM / flow)
→ decoder → audio 44.1 / 48 kHz
best for loops, textures, sound design
weaker structure, weaker vocals
2026 leaderboard + legal note
MusicGen-large 3.3B 30 s no vocals MIT - instrumental baseline
Stable Audio Open 1.2B 47 s no vocals non-commercial - loops, textures
ACE-Step XL (Apr 26) 4B > 2 min vocals Apache-2.0 - open full-song
Suno v5 (closed) ? 4 min vocals commercial - ELO 1293; quality leader
Udio v4 (closed) ? 4 min vocals+stems commercial - inpainting, stem split
legal: Warner + UMG settled with Suno / Udio in 2025-26; disclosure + watermarking required (EU AI Act, CA SB 942)
safe-to-ship: instrumental MIT/CC0 · commercial API with license · train on owned catalog