MetaVoice-1B

MetaVoice-1B is a 1.2-billion-parameter text-to-speech base model released in February 2024 by the MetaVoice startup (unrelated to Meta). Trained on 100,000 hours of speech, it prioritizes emotional speech rhythm and tone in English, zero-shot cloning of American and British voices from 30 seconds of reference audio, and long-form synthesis.

Released under the Apache 2.0 license with a reference implementation on GitHub, it combines causal and non-causal transformers with multi-band diffusion and deep filter networks, and was designed to avoid generating non-existent words, a common failure mode in competing models. Date is era-level (February 2024).