Raon-OpenTTS: an open TTS model for English only

More than the weights: the entire TTS stack is open

What stands out here is not simply another TTS model, but a rare attempt to release the full stack: weights, training data, training code and inference code. In the Raon-OpenTTS repository and model card, KRAFTON AI describes the system as open-data and open-weight, built for zero-shot synthesis from reference audio. As of August 21, 2026, it is more accurate to view this as an analysis of a published project rather than breaking news, since the supplied materials do not state the date of the original announcement.

Two variants are available. Raon-OpenTTS-0.3B has 336M parameters, while Raon-OpenTTS-1B has 1048M. Both use a DiT architecture derived from F5-TTS, making the choice fairly practical: the smaller model suits more constrained environments, while the larger one is for testing a potential quality gain.

The most substantial part of the release sits beneath the model itself. The Raon-OpenTTS-Core corpus contains 510.1K hours after filtering an original pool of 615K hours drawn from 11 English-language datasets. The filtering pipeline, training code and checkpoints are public as well, making the project compelling for reproducible experiments rather than merely running ready-made inference.

On the Seed-TTS-Eval benchmark, the smaller model achieved WER 1.95 and SIM 0.687, while the larger model reached WER 1.78 and SIM 0.749. A separate Raon-OpenTTS-Eval set focuses on robustness: 6,000 reference-text pairs, four acoustic conditions and 12 datasets. These are the authors' published results, not my own measurements, but the direction is clear: scaling up the model improves both reported metrics.

What actually changes for developers

The primary benefit is not a flashy demo but the ability to inspect how the result was produced. The open corpus and training pipeline make it possible to study filtering, reproduce training and adapt individual parts of the system without depending on a closed API.

For voice pipelines, this provides a clear zero-shot TTS foundation based on reference audio. Still, the published benchmarks do not answer questions about latency, memory use or robustness for specific voices. I would test those properties first, and separately for both model sizes: the gap between an impressive table and usable inference usually becomes visible there.

The limitation is strict: training used English speech only. Russian and other languages fall outside the documented use case, so occasional pronunciation of isolated phrases should not be treated as multilingual support. Raon-OpenTTS therefore looks like a strong open foundation for English-only TTS, but it does not yet replace a universal voice layer.

We previously covered Rust LocalGPT, a local assistant designed for practical self-hosted AI use. Raon-OpenTTS similarly expands the options for teams that want more control over their AI stack.