Technical Inquiry: Generating Traditional Chinese Yue Opera with YuE

[oooranger817]

Hello YuE Team and Community,

First, thank you for this incredible open-source model!

I’m a student working on a project exploring AI-generated music for 'Yue Opera (越剧)', a traditional Chinese opera. I was deeply inspired by the “Shaanbei Folk Song” demo and the description of the model‘s encoding of diverse musical heritage.

I have several technical questions and would be very grateful for any guidance:

1. Training Data: Was 'Yue Opera' or similar traditional Chinese opera audio data included in the training corpus? If so, what was the approximate scale or mixing ratio?
2. Prompt Engineering: To guide the model towards the Yue Opera style, what key genre tags would you recommend? Can the tag combination from the “Shaanbei Folk Song” demo serve as a reference?
3. Fine-tuning Path: Given the support for LoRA, if the base model‘s exposure to Yue Opera is limited, is LoRA fine-tuning with a dedicated dataset the most promising path for high-quality generation? Any practical advice on this?
4. Effectiveness of Audio ICL: How effective is the dual-track ICL mode in imitating the specific vocal style and instrumentation of a provided Yue Opera clip? Is it significantly better than CoT (text-only) guidance for such concrete styles?
5. Inference Parameters: For generating a complete 3-4 minute Yue Opera piece, are there any critical inference parameters (e.g., --repetition_penalty, --run_n_segments) that need special adjustment?

Any insights would be invaluable for our cultural heritage exploration project. Thank you for your time and contribution!