New state-of-the-art speech language models are here 🚀
Meet Raon, KRAFTON AI’s open model series.
🎙️ Raon-SpeechChat [🎮 Demo: raon.krafton.ai]
A full-duplex speech model that listens and speaks at the same time
Backchannel. Interrupt. Respond. Overlap.
Just like a real conversation.
Best average performance across key conversational behaviors.
🧠 Built on Raon-Speech (9B) [🎮 Demo: raon.krafton.ai]
A State-of-the-art bilingual (EN/KO) 9B speech language model.
We also release
🔊 Raon-OpenTTS (0.3B released, 1B TBD)
Small TTS models trained on fully open datasets with their full pipeline.
👁️ Raon-VisionEncoder (400M)
Trained entirely on public data from scratch.
All model checkpoints, training/inference pipelines, and an interactive demo are open-sourced (OpenTTS includes datasets).
Built to be extremely easy to use. Ready to run out of the box.
🤗 HuggingFace: huggingface.co/collections/KR…
📄 Technical Report (Speech & SpeechChat): huggingface.co/KRAFTON/Raon-S…
📝 Blog (VisionEncoder): krafton.ai/blog/posts/rao…
