Kakao Enhances Kanana-o with LM-SPT for Efficient, Style-Controllable Voice Generation

Kakao integrated LM-SPT, a low-frame-rate speech tokenizer, into Kanana-o to enhance voice generation efficiency and style control. LM-SPT operates at 12.5 frames per second, simplifying the architecture and reducing computational load. The system uses Multi-objective online reinforcement learning to better follow natural language style instructions. On the Korean InstructTTSEval benchmark, Kanana-o achieved a score of 94.50, demonstrating performance comparable to, though slightly below, Gemini-2.5-flash-preview-tts (95.38). Additionally, the model demonstrated cross-language transfer, applying style controls learned in Korean to English speech generation.