MetaVoice is a foundational text-to-speech model family built for expressive, human-like speech. Its open-source release, MetaVoice-1B, synthesizes speech with emotion and rhythm, and can clone a voice from as little as 30 seconds of reference audio without fine-tuning. The project has about 4,200 stars on GitHub.
Most TTS systems sound flat because they optimize for intelligibility; MetaVoice was trained to capture the texture of real voices — pauses, emphasis, laughter-adjacent expression — using a causal, streaming-friendly architecture that generates audio token by token. The repository contains the model code and weights for the 1B-parameter model, supporting English plus several other languages, with zero-shot voice cloning driven by a short sample and text prompts. It was released by the MetaVoice startup (unrelated to Meta Platforms) as the open core of its voice platform, and it influenced a wave of open TTS work on expressive synthesis. The repository was last updated in 2024, so treat it as a solid, finished research release rather than a fast-moving project; the weights remain a strong self-hosted option for character voices, narration and prototyping voice products.
Technically, MetaVoice combines a text-to-semantic model with an acoustic model that renders fine waveform detail, trained with speaker and emotion conditioning so a reference clip transfers not just timbre but speaking style. The 1B model was trained on large-scale English audiobook and podcast-style data, and the repository includes guidance on sampling settings for stability. Note the naming: this MetaVoice is an independent voice-AI startup, and the open release is a snapshot of its earlier research model rather than its current hosted platform.
The open-source model is free under the Apache-2.0 license; download the weights and run them on your own hardware.
MetaVoice is for developers and creators who need expressive, clonable voices for games, videos, audiobooks or voice products and want an open model they can self-host.