Krafton Open-Sources 21B Speech-to-Speech Voice AI Model, Bringing Real-Time Conversational Game Characters to Life
Technology📅 July 29, 2026👤 FreeReadText Team

Krafton Open-Sources 21B Speech-to-Speech Voice AI Model, Bringing Real-Time Conversational Game Characters to Life

South Korean gaming giant Krafton releases A.X K2 Raon-Speech on Hugging Face, a 21-billion-parameter end-to-end speech-to-speech model that ranks #1 in Korean and #3 in English among open-source voice AI — designed to power game characters that understand player voice and intent in real time.

On July 29, 2026, South Korean gaming company Krafton — best known for PUBG: Battlegrounds — released A.X K2 Raon-Speech on Hugging Face, a 21-billion-parameter speech language model that converts speech directly to speech without passing through an intermediate text representation. The model builds on SK Telecom's A.X K2 small language model, paired with Krafton's in-house speech encoder and codec, creating an end-to-end architecture that preserves the speaker's emotion, tone, and intonation — qualities that traditional cascaded ASR-to-text-to-TTS pipelines systematically lose.

The benchmark results position Raon-Speech as a serious contender in open-source voice AI. Among publicly available speech language models with 30 billion parameters or fewer, it ranks #1 overall in Korean and #3 in English, evaluated across 46 benchmarks spanning six domains: speech recognition, speech synthesis, speech understanding, spoken question answering, text question answering, and tool calling. The evaluation suite includes 24 Korean benchmarks and 22 English benchmarks, covering standard datasets like LibriSpeech, KSponSpeech, VoiceBench, and FunctionChat-Bench. The model's strength in Korean is particularly notable given the relative scarcity of high-quality open-source voice AI for non-English languages.

The model is part of Krafton's 'Raon' AI brand, launched in April 2026 with four initial open-source models: Raon-Speech (voice language), Raon-SpeechChat (real-time voice conversation), Raon-OpenTTS (text-to-speech), and Raon-VisionEncoder (vision). A.X K2 Raon-Speech represents a second-generation upgrade over the original Raon-Speech, using Krafton's self-trained encoder and codec for significantly improved performance. The project is also part of South Korea's government-backed 'Sovereign AI Foundation Model' initiative, overseen by the Ministry of Science and ICT, with Krafton participating as part of an SK Telecom-led elite research consortium — reflecting a broader national strategy to develop independent AI capabilities.

Krafton's ultimate application is its Co-Playable Character (CPC) system: AI companions within games that can understand a player's voice and intent in real time and respond with situation-appropriate, emotionally expressive speech — moving far beyond the fixed, pre-scripted dialogue trees that have defined game NPCs for decades. CEO Kim Chang-han has positioned AI as central to Krafton's future, and Chief AI Officer Lee Kang-wook described Raon-Speech as 'the result of Krafton's consistent research and advancement in AI foundation technologies.' For the broader voice AI industry, Krafton's entry signals that gaming — with its real-time interactivity requirements, massive user bases, and willingness to invest in novel AI — could become a primary driver of speech-to-speech technology alongside call centers and voice assistants.

KraftonRaon-SpeechGaming AISpeech-to-SpeechOpen SourceKorean AIGame Characters

出典

← Back to News