Energy-based models (EBMs) are appealing for text-to-speech (TTS) because they can represent flexible distributions over acoustic outputs, but they come with a catch: the normalisation term is intractable. Noise contrastive estimation (NCE) has been a popular workaround, training the model by comparing unnormalised log-likelihoods. However, NCE has known drawbacks, and the new arXiv paper argues for a different route.
The paper, titled "Score-Based Training for Energy-Based TTS Models," proposes a score-based training objective that avoids the need for explicit normalisation altogether. Instead of contrasting against noise samples as NCE does, the method learns the score—the gradient of the log-density—which is also normalisation-free. This shift could lead to more stable training or better sample quality, though the abstract does not yet provide experimental details.
For researchers working on generative models for speech, this is a conceptual contribution that adds to the toolbox for training EBMs. It directly addresses a known weakness of NCE and suggests a principled alternative. The full paper, available on arXiv, presumably contains the derivation and any empirical validation, but the core idea is clear from the abstract alone.