Krafton said on Tuesday it has released its voice artificial intelligence (AI) foundation model, 'A.X K2 Raon-Speech', on global AI platform Hugging Face.

The model is a second-phase research outcome by Krafton, which is participating as part of an elite SK Telecom team in a Ministry of Science and ICT-led project on proprietary AI foundation models. It combines SK Telecom’s small language model (A.X K2) with Krafton’s in-house voice encoder and codec to directly understand and generate speech.

The key is a system that converts speech directly between voices. It overcame a limitation in which emotion and intonation in a user’s voice were lost when existing AI converted speech to text for understanding and then synthesized it back into speech.

Krafton trained in-house a voice encoder that converts human voices into AI information and a voice codec that restores it into natural speech. It was designed to reproduce a person’s voice tone and emotion as-is by stably linking voice input and output while maintaining the language model’s text-processing performance.

With 21 billion parameters, 'A.X K2 Raon-Speech' ranked No. 1 in overall Korean performance and No. 3 in English among publicly available speech language models of 30 billion parameters or less. It went through 46 benchmark evaluations across six areas, including speech recognition, speech synthesis, speech understanding, speech question-and-answer, text question-and-answer and tool calling, with 24 in Korean and 22 in English.

Krafton previously launched its AI brand 'Raon' in April and released four types as open source, including a speech language model, a real-time voice conversation model, a text-to-speech conversion model and a vision encoder. The new model is an advanced result based on those underlying technologies.

Kangwook Lee (이강욱), Krafton’s chief AI officer (CAIO), said, "A.X K2 Raon-Speech is the result of Krafton’s steady research and advancement of AI foundation technology." He added, "We will continue to advance AI foundation technologies needed for games to create new game experiences."

Krafton plans to apply the voice AI technology accumulated through this research to advance its in-game companion characters called CPC (Co-Playable Character). It aims to build an environment in which AI characters, beyond outputting preset lines, understand a user’s voice and intent in real time and interact with situation-appropriate speech.