Subscribe

Voicebox

Voicebox is a non-autoregressive flow-matching model trained on 50K+ hours of speech to infill audio, covering zero-shot TTS, noise removal, editing and style transfer.

Generalist speech model that learns in-context from audio plus text, usable across languages. Meta published the paper and demos without releasing the model.

Date
Friday 23 June 2023
Lab
Meta (FAIR)
Kind
paper
Access
research preview

The date is the arXiv first version, posted 23 June 2023.

Sources

  1. arxiv.org/abs/2306.15687

This record was checked against its sources on 6 October 2026. How we check