Voicebox

Voicebox is a generative artificial intelligence model developed by Meta, the parent company of Facebook. The model was announced in 2023 as a massive breakthrough in AI-driven audio and speech synthesis. Meta designed it to perform tasks it was not explicitly trained to do, utilizing a highly versatile architecture.

Voicebox utilizes a non-autoregressive flow-matching architecture, allowing it to generate audio much faster than traditional autoregressive models. It is capable of zero-shot text-to-speech generation, meaning it can clone a voice from just a two-second audio sample. This architecture also allows the model to perform tasks like noise removal and cross-lingual speech generation.

When Meta announced Voicebox in 2023, they demonstrated its ability to generate speech in multiple languages, including English, French, Spanish, German, and Italian. The model can take an audio sample in one language and generate text-to-speech in another while preserving the original speaker's voice. This makes it a powerful tool for global translation services.