Reproduce “Prompt-Singer: Controllable Singing-Voice-Synthesis with Natural Language Prompt” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yongqi Wang, Ruofan Hu, Rongjie Huang, Zhiqing Hong, Ruiqi Li, Wenrui Liu, Fuming You, Tao Jin and 1 more

Prompt-Singer introduces a controllable singing-voice-synthesis method that allows users to adjust singer gender, vocal range, and volume using natural language prompts. It achieves attribute control accuracies up to 94.9% while maintaining audio quality comparable to non-controllable SVS baselines.

Key insights

  • Decoupling fundamental frequency (F0) into an average F0 (vocal range factor) and a rescaled F0 (melodic contour) enables precise vocal range and gender control while preserving the original melody.
  • Fine-tuning pre-trained text encoders on specific style labels significantly improves the model's ability to interpret and apply natural language prompts for singing voice synthesis.
  • Incorporating speech data during training effectively alleviates data scarcity in singing voice synthesis, leading to improved controlling accuracy and generation quality.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.