New energy-based models offer alternative for efficient speech synthesis

Researchers have introduced a new type of non-autoregressive model for speech synthesis, known as energy-based models (EBMs). This approach utilizes noise contrastive estimation for training and explores various strategies for generating effective negative samples, including the use of high-performing autoregressive models. The paper also details how Langevin Markov Chain Monte-Carlo (MCMC) can be employed for sampling from EBMs, drawing parallels between EBMs and diffusion models. Experimental results on the LJSpeech dataset indicate that this novel EBM approach outperforms Tacotron 2. AI

IMPACT Introduces a novel energy-based model architecture for speech synthesis, potentially improving inference efficiency compared to autoregressive models.

RANK_REASON The cluster describes a research paper introducing a new model architecture for speech synthesis. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• FastSpeech 2
• LJSpeech
• Tacotron 2
• Wanli Sun

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →