Reproduce “MM-TTS: Multi-modal Prompt based Style Transfer for Expressive Text-to-Speech Synthesis” 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
Wenhao Guan, Yishuang Li, Tao Li, Hukai Huang, Feng Wang, Jiayan Lin, Lingyan Huang, Lin Li and 1 more
MM-TTS presents a framework for expressive Text-to-Speech synthesis that unifies style control across multi-modal prompts, including reference speech, emotional facial images, and text descriptions. This system yields improved naturalness and style similarity in synthesized speech when compared to single-modality baselines.
Key insights
- Pre-trained multi-modal models like CLIP can be effectively adapted to align diverse input modalities (text, image, speech) into a unified style latent space for TTS.
- Dynamic prediction of convolution kernels and biases (SAConv) facilitates more localized and accurate style transfer compared to global adaptive normalization techniques.
- Rectified Flows offer an efficient and high-quality method for refining Mel-spectrograms, addressing over-smoothing with significantly fewer inference steps than diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.