[Edaily Reporter An Yu-ri ] KRAFTON(259960)announced on the 29th that it has released its voice AI foundation model, “A.X K2 Raon-Speech,” on the global AI platform Hugging Face. With 21 billion parameters, the model ranked first in overall Korean language performance among publicly available speech models with 30 billion parameters or fewer.

KRAFTON is leading research on next-generation multimodal models as part of SKTelecom’s elite team in the “Independent AI Foundation Model” project, organized by the Ministry of Science and ICT. A.X K2 Raon-Speech is a speech language model released as one of the outcomes of the project’s second phase.

KRAFTON emphasized that the model, with 21 billion parameters, ranked first in overall performance for Korean and third in overall performance for English among publicly available speech-language models with 30 billion parameters or fewer.

The evaluation was conducted across a total of six domains: speech recognition, speech synthesis, speech understanding, speech question-answering, text question-answering, and tool invocation. A total of 46 benchmarks were utilized, including 24 for Korean and 22 for English, such as LibriSpeech, KSponSpeech, VoiceBench, KVoiceBench, MMAU, KMMAU, API-Bank, and FunctionChat-Bench.

A.X K2 Raon-Speech is a speech language model trained to understand and generate speech by applying KRAFTON’s self-trained speech encoder and codec to SKTelecom’s A.X K2-based compact language model. KRAFTON independently trained a speech encoder that converts human voices into information understandable by AI, as well as a speech codec that restores AI-generated information into natural-sounding speech. The company developed and applied a training method that reliably links speech input and output while maintaining the language model’s text capabilities.

AI speech dialogue systems typically operate by converting the user’s speech into text to understand it, generating a text-based response, and then synthesizing that response back into speech. During this process, information contained in the user’s voice—such as emotions and intonation—can be lost. A.X K2 Raon-Speech is a model capable of generating natural-sounding voice responses by leveraging the information embedded in the user’s voice.

Last April, KRAFTON launched its AI model brand “Raon” and released four AI models as open source: a speech language model (Raon-Speech), a real-time voice dialogue model (Raon-SpeechChat), a text-to-speech model (Raon-OpenTTS), and a vision encoder (Raon-VisionEncoder).

The newly unveiled A.X K2 Raon-Speech represents a technological advancement achieved by applying KRAFTON’s proprietary voice encoder and codec—trained in-house—to SKTelecom’s compact language model, building upon the training methodology of the speech language model (Raon-Speech) released at that time.

KRAFTON plans to leverage the speech AI capabilities accumulated through this research to further refine CPCs (Co-Playable Characters), which are AI characters that play alongside users. The company aims to create a gaming experience where users and AI interact more naturally, with CPCs understanding users’ voices and intentions and responding with speech tailored to the game situation.

Lee Kang-wook, CAIO at KRAFTON, stated, “A.X K2 Raon-Speech is the result of KRAFTON’s consistent research and refinement of AI foundation technologies,” adding, “We will continue to advance the AI foundation technologies necessary for gaming to create new gaming experiences.”