SK Telecom elite team's multimodal research yields second-phase results

Model reads intonation in voice to generate natural speech responses

Krafton's headquarters building. [Krafton]
Krafton's headquarters building. [Krafton]

Krafton has unveiled a voice AI foundation model called A.X K2 Raon-Speech.

Krafton released A.X K2 Raon-Speech on Wednesday on Hugging Face, a global AI platform. The model is the result of research the company conducted as part of an elite SK Telecom team participating in the government's independent AI foundation model project, where it worked on next-generation multimodal model research.

Krafton described A.X K2 Raon-Speech as a voice language model released as one of the project's second-phase achievements. The model has 21 billion parameters.

Krafton said the model ranked first in overall Korean-language performance and third in overall English-language performance among publicly available voice language models with 30 billion parameters or fewer. The evaluation covered six areas — speech recognition, speech synthesis, speech understanding, spoken question answering, text-based question answering and tool calling — using a total of 46 benchmarks: 24 in Korean and 22 in English.

A.X K2 Raon-Speech was built by applying Krafton's own trained voice encoder and codec to SK Telecom's A.X K2-based small language model, producing a voice language model capable of both understanding and generating speech.

Krafton said it independently trained a voice encoder that converts human speech into information an AI can process, and a voice codec that reconstructs AI-generated information into natural-sounding speech. The company also developed and applied a training method that stably connects voice input and output while preserving the language model's text capabilities.

According to Krafton, conventional AI voice conversation systems have struggled to retain information embedded in a user's voice — such as emotion and intonation. They typically work by converting speech to text, generating a text response and then synthesizing that response back into speech. A.X K2 Raon-Speech, by contrast, can draw on information carried in the user's actual voice to produce more natural spoken responses.

In April, Krafton launched an AI model brand called "Raon" and released four open-source AI models: a voice language model (Raon-Speech), a real-time voice conversation model (Raon-SpeechChat), a text-to-speech model (Raon-OpenTTS) and a vision encoder (Raon-VisionEncoder).

The new A.X K2 Raon-Speech builds on the training approach behind the Raon-Speech model released at that time, applying Krafton's own trained voice encoder and codec to SK Telecom's small language model to advance the technology a step further.

"A.X K2 Raon-Speech is the result of Krafton's consistent research and refinement of AI foundation technology," said Lee Kang-wook, Krafton's chief AI officer. "We will continue to advance the AI foundation technology needed for games and create new gaming experiences."


chami@heraldcorp.com