- Home
- Text-to-speech tools
- MetaVoice-1B
On the recordRead 13 Oct 202624 facts
MetaVoice-1B
On the record
- Access options
- The page lists local use, Hugging Face, and a Google Colab demo as ways to use the model.github.com · read 12 Oct 2026
- API
- The inference server provides API definitions at its /docs route.github.com · read 13 Oct 2026
- API access
- Yesgithub.com · read 20 Sep 2026
- Commercial use
- Yesgithub.com · read 20 Sep 2026
- Dataset caveat
- The fine-tuning workflow does not check for overlap between training and validation datasets, so the page advises keeping them disjoint.github.com · read 12 Oct 2026
- Deployment
- The project provides a web UI and inference server that can be deployed on AWS, GCP, or Azure.github.com · read 12 Oct 2026
- Expressive speech
- The project prioritizes emotional speech rhythm and tone in English.github.com · read 12 Oct 2026
- Fine-tuning
- The project supports fine-tuning the first-stage language model and describes cross-lingual voice cloning with fine-tuning.github.com · read 13 Oct 2026
- Fine-tuning data limit
- The fine-tuning instructions say the project does not check for overlap between training and validation datasets.github.com · read 13 Oct 2026
- Hardware requirement
- Installation requires a GPU with at least 12 GB of VRAM, Python 3.10 to below 3.12, and pipx.github.com · read 12 Oct 2026
- Hardware requirements
- Installation prerequisites include a GPU with at least 12 GB of VRAM and Python 3.10 or later but earlier than 3.12.github.com · read 13 Oct 2026
- Integrations
- The README describes an optional Weights & Biases integration for logging fine-tuning runs.github.com · read 13 Oct 2026
- License
- The model is released under the Apache 2.0 license, which the project describes as usable without restrictions.github.com · read 13 Oct 2026
- Performance
- After compilation, the synthesise API runs faster than real time on Ampere, Ada-Lovelace, and Hopper GPUs, according to the project page.github.com · read 12 Oct 2026
- Purpose
- MetaVoice-1B is a 1.2-billion-parameter text-to-speech model trained on 100,000 hours of speech.github.com · read 13 Oct 2026
- Quantization limit
- Experimental int4 and int8 inference can run faster, but the project says this degrades audio quality.github.com · read 13 Oct 2026
- Speech style
- The model is designed for emotional speech rhythm and tone in English.github.com · read 13 Oct 2026
- Startup time
- The reference inference script takes 30 to 90 seconds to start, depending on hardware, because it compiles the model.github.com · read 12 Oct 2026
