Practical overview
When it fits
Use MetaVoice-1B when you want an open-source model and documented implementation for text-to-speech experimentation, especially where expressive delivery or voice cloning is relevant.
Who it is for
- Developers building text-to-speech applications or expressive speech-generation prototypes.
- Researchers exploring speech synthesis, zero-shot voice cloning or first-stage model fine-tuning.
- Teams able to provide the documented GPU and Python environment for local inference.
Documented uses
- Generate spoken audio from supplied text with the reference implementation.
- Prototype voice cloning using American or British voice reference audio.
- Fine-tune the first-stage language model using paired audio and caption paths.
Requirements & limits
The repository specifies an Apache-2.0 license and links a Colab demo and hosted playground. Local setup requires a GPU with at least 12 GB of VRAM and Python >=3.10,
- The documented local prerequisites include a GPU with at least 12 GB of VRAM.
- The model's first inference startup takes 30–90 seconds depending on hardware because it compiles the model.
- The README says quantized inference is experimental and may reduce audio quality.
- Fine-tuning requires users to ensure that training and validation data do not overlap.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up dependencies for local MetaVoice-1B use
Prepare the documented environment and install the project dependencies for a local inference workflow.
Before you start
- GPU with at least 12 GB of VRAM
- Python >=3.10,<3.12
- pipx
- ffmpeg and Rust
Steps
Install ffmpeg and Rust as described in the repository README.
Install the Poetry dependency manager with pipx.
Install project dependencies with Poetry, then install torch==2.2.1 and torchaudio==2.2.1.
Expected outcome
The README's documented environment and project dependencies are installed for the subsequent inference stage.
Guide limitations
- This guide covers setup only; it does not verify successful installation or inference.
- The README also documents a pip/conda dependency-installation route.
Prepare fine-tuning data and run the documented training command
Use the repository's sample fine-tuning workflow with a pipe-delimited dataset.
Before you start
- The project dependencies installed using the repository instructions
- Training and validation CSV files with audio_files and captions columns
- Training and validation datasets that do not overlap
Steps
Format each dataset as a pipe-delimited CSV with audio_files and captions fields.
Keep the training and validation datasets disjoint.
Run the documented finetune command with the training and validation CSV paths.
Use the resulting first-stage checkpoint with the documented inference command.
Expected outcome
The documented workflow trains a first-stage model and specifies how to pass its checkpoint to inference.
Guide limitations
- The README says the project does not perform dataset-overlap checks.
- This recipe reflects documented commands and is not a verification of training or inference.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- metavoiceio/metavoice-src · Official homepage title · retained in its source language