The model, announced by Krafton on Wednesday, was developed under a government-led "sovereign AI foundation model" project overseen by the Ministry of Science and ICT, in which Krafton leads next-generation multimodal research as part of an elite team assembled by SK Telecom.
It emerged as one of the results of the project's second phase.
Built with about 21 billion parameters, A.X K2 Raon-Speech ranked first in overall Korean-language performance and third in English among openly released voice models below the 30-billion-parameter class, according to Krafton.
The company said it evaluated the model across six areas — from speech recognition and synthesis to tool-calling — using 46 benchmarks spanning Korean and English.
The model pairs a compact language model based on SK Telecom's A.X K2 with a speech encoder and codec that Krafton trained in-house, allowing it to interpret a speaker's voice and answer in fluent, natural-sounding speech.
Conventional voice systems, which convert speech to text and back again, often strip away emotion and intonation along the way; Krafton said its model preserves those cues to craft more lifelike replies.
The release builds on Raon, the AI brand Krafton launched in April with four open-source models, refining the earlier speech model's training approach with the company's own encoder and codec.
"A.X K2 Raon-Speech is the fruit of Krafton's steady research into and refinement of AI foundation technology," said Lee Kang-wook, the company's chief AI officer, adding that the firm would keep advancing the foundations needed to build new gaming experiences, including AI-driven characters that understand a player's voice and respond in kind.
Copyright ⓒ Aju Press All rights reserved.
