Krafton on Wednesday released its new audio foundation model, A.X K2 Raon-Speech, on global AI platform Hugging Face.
The company is part of SK Telecom’s consortium participating in the Ministry of Science and ICT’s Independent AI Foundation Model project, where it leads research on next-generation multimodal models.
A.X K2 Raon-Speech builds on Krafton’s existing Raon-Speech model. It combines the company’s proprietary voice encoder and audio codec with a compact language model based on SK Telecom’s A.X K2, allowing the system to process and generate speech directly.
The model has 21 billion parameters and ranked first among Korean models and third among English-language models with fewer than 30 billion parameters, according to Krafton.
The evaluation covered six areas: speech recognition, speech synthesis, speech understanding, spoken question answering, text-based question answering and tool use. It used 24 Korean and 22 English benchmarks, including LibriSpeech, KSponSpeech and VoiceBench.
Many conversational audio systems first convert speech into text, generate a written response and then turn it back into audio. Krafton said this process can lose information carried through emotion, tone and intonation.
A.X K2 Raon-Speech is designed to retain those cues and produce more natural, humanlike voice responses.
Krafton plans to apply the technology to its AI-powered co-playable characters, for more natural conversation and responses during gameplay.
The release follows four open-source models introduced under Krafton’s Raon brand in April: Raon-Speech, Raon-SpeechChat, Raon-OpenTTS and Raon-VisionEncoder.
“A.X K2 Raon-Speech reflects Krafton’s continued research and refinement of core AI technologies,” said Lee Kang-wook, chief AI officer at Krafton. “We will continue advancing our foundation models to create new gaming experiences.”
stlee0329@heraldcorp.com