VoiceBox

VoiceBox, announced by Meta on June 16, 2023, is a text-guided generative model for speech built on flow matching rather than autoregression. Trained on over 50,000 hours of unfiltered public-domain audiobook speech across six languages, it learns a text-guided speech infilling task: predicting masked audio segments from surrounding context and transcripts.

That single training objective generalizes through in-context learning to zero-shot TTS, noise removal, content editing, and style conversion, reportedly beating VALL-E on intelligibility and audio similarity while running up to 20x faster. Meta deliberately did not release the model or code, citing misuse risk, so VoiceBox is a research landmark rather than a usable tool; its techniques live on in later open speech models.