arXiv · 2023 · Preprint
Tong et al. (Mila - Québec AI Institute) · → Paper · Demo: ? · Code: ✓
Introduces conditional flow matching (CFM) and its optimal transport variant (OT-CFM), simulation-free training objectives for continuous normalizing flows that remove the Gaussian source requirement, stabilise training, and enable faster inference through straighter transport trajectories.
Citation Stub
This paper is not a speech generation paper but is cited by the corpus. See Context in Speech Generation below for why it is relevant.
Context in Speech Generation
This paper introduces the conditional flow matching (CFM) framework as a unifying formulation for training continuous normalizing flows (CNFs) without simulation. The key variant, OT-CFM, pairs source and target samples according to a minibatch optimal transport plan rather than an independent coupling, producing straighter probability paths that converge in fewer neural network evaluations during inference. The framework’s simulation-free regression objective, which avoids integrating the ODE during training, makes CNF-based generative models substantially more practical to train at scale. TTS systems can adopt flow matching as the core generative mechanism for mel-spectrogram or latent acoustic feature synthesis, benefiting from deterministic inference, flexible source distributions, and the speed-quality trade-off that straighter OT flows provide.
Wiki Connections
- flow-matching — This paper introduces the conditional flow matching framework and the OT-CFM variant that underlies flow-matching-based TTS systems requiring efficient, simulation-free training of continuous normalizing flows.