Best Practices for Building Suno-Style Music Generation Models: What Works?

20 Eki 03:45

I'm exploring generative models for music synthesis inspired by Suno's approach. I have a modest dataset of MIDI and audio stems and I'm considering transformer‑based or diffusion architectures. What preprocessing pipelines (e.g., tokenization, tempo alignment) work best for preserving musical structure? How do you handle long‑range dependencies without blowing up memory? Any tips on evaluation metrics beyond subjective listening—perhaps using pitch‑accuracy or rhythmic consistency scores? Would you recommend multi‑modal conditioning (lyrics, genre tags) at training time? Looking forward to your experiences.