Voicebox is a machine learning model that generates speech from text. It can be used to customize voices, power applications and help non-verbal individuals communicate better. Meta's unique training method and use of text-guided speech infilling make it highly efficient and generalizable for varied speech data.