Kakao's Kanana-o AI: Advanced Voice Control & Emotions

Kakao has significantly advanced the voice generation function of its AI model, "Kanana-o." This updated model can now control speech, emotions, and accents in detail. What's interesting is that you can instruct the AI using natural language, like "Read it very quickly" or "Read it in Gyeongsang-do dialect." The AI will then reflect speed, volume, pitch, emotions, intonations, and intensity. It can even take role-based instructions, such as "Read it like a sports broadcast." While trained primarily on Korean data, it can apply these same speech instructions to English voice generation. Kanana-o scored 94.50 points in the Korean language category of the InstructTTSEval benchmark, surpassing OpenAI's GPT-4o-mini-tts. Kakao also improved generation speed and efficiency using its own voice tokenizer, 'LM-SPT'. The company plans to add non-verbal expressions like laughter and sighs in the future. The bottom line is that Kakao aims to provide a more natural and convenient AI voice experience in various services.