KRAFTON/ax-k2-raon-speech-demo

A.X K2 Raon-Speech · 21B-A3B

Demo of `KRAFTON/A.X-K2-Raon-Speech-21B-A3B`, a bilingual English/Korean speech language model (21.2B total / 3.5B active parameters) built on SK Telecom's A.X K2 Light MoE backbone with KRAFTON AI's AuT speech encoder and Mimi-style neural audio codec.

Three tasks, one model:

• —Transcribe — speech → text (STT)
• —Ask about audio — SpeechQA (audio + text question) and SpokenQA (answer the question spoken inside the clip)
• —Speak — text → speech, optionally cloning a reference speaker's voice or continuing a reference recording's prosody

Inference follows the official reference implementation (krafton-ai/Raon-Speech `demo/gradio_demo.py`) via the model's own RaonPipeline, with the task defaults from the upstream config/infer.yaml.

Example audio comes from the upstream repo's data/speechllm/eval samples (LibriTTS-R / LibriHeavy / SLUE / AI-Hub, permissively licensed).

Model license: CC BY-NC 4.0 — research and non-commercial use only.