For years, speech generation lagged text and image generation in both scale and generality — every task got its own bespoke model. Meta AI's Voicebox broke that pattern. It's a non-autoregressive flow-matching model trained on over 50,000 hours of unfiltered speech to do one thing well: infill audio given surrounding context and text. That single objective unlocks zero-shot TTS, cross-lingual synthesis, noise removal, content editing, and style transfer, all from the same checkpoint.

The review walks through why flow matching (and the neural ODE machinery underneath it) is such a good fit for speech — continuous, non-autoregressive, and dramatically faster than diffusion at comparable quality. It also puts the headline numbers in context: Voicebox beats VALL-E on both intelligibility (1.9% vs 5.9% WER) and speaker similarity while running up to 20x faster. If you're tracking where speech foundation models are headed after VALL-E and Whisper, or you want to actually understand what flow matching is doing instead of just reading the buzzword, the deep-dive is worth a listen.