Kakao upgrades Kanana-o AI to express emotions, accents via text prompts

An infographic of Kakao's self-developed omni AI model, Kanana-o's performance / Courtesy of Kakao

Kakao, the operator of Korea’s dominant messaging app KakaoTalk and a major digital services ecosystem, has upgraded the voice generation technology in its proprietary omni artificial intelligence (AI) model, Kanana-o. The update allows users to fine-tune tone, speed, emotions and accents through simple natural-language instructions alone.

According to a post on Kakao's tech blog published Tuesday, the improved model lets users shape how speech is delivered — not just what is read aloud — by giving instructions such as "read it very quickly," "read it in a low voice," "read it in a sad voice" or "read it in a Gyeongsang Province accent."

The AI adjusts speed, volume, pitch, emotion, intonation and intensity accordingly.

The model can also follow role-based instructions, such as reading "like a sports broadcast," "like a news anchor" or "like a storybook," and can combine multiple instructions at once — for example, "lower the tone and read it quickly in a sad voice."

Kakao said that although the model was trained primarily on Korean-language data, it can follow the same kinds of instructions in English as well.

On the Korean-language "InstructTTSEval" benchmark, which measures how well a model follows spoken-delivery instructions, Kanana-o scored 94.50 points — ahead of OpenAI's GPT-4o-mini-tts, which scored 91.10, and close to Google's Gemini-2.5-flash-preview-tts, which scored 95.38.

Kakao also said it improved the speed and efficiency of voice generation by applying a newly developed in-house voice tokenizer called LM-SPT (LM-aligned SPeech Tokenizer), which compresses speech into fewer tokens, reducing the amount of data the AI needs to process.

In evaluations comparing Korean- and English-language voice understanding and generation across various speech language models, LM-SPT outperformed recent global technologies including Mimi, DualCodec and CosyVoice2, and scored highest in expert evaluations of the naturalness and speaker similarity of synthesized speech.

Kakao said it plans to continue developing Kanana-o's voice technology, including research into processing voice understanding and generation within a single, integrated framework to deliver more natural, seamless speech, as well as work on generating non-verbal expressions such as laughter, sighs and exclamations for finer voice control.

"This upgrade to Kanana-o's voice technology was focused on generating speech that sounds as natural as a human voice, while giving the model the ability to express whatever tone, emotion or accent users want through natural-language instructions," a Kakao official said.

"We plan to apply the Kanana-o model to a range of services going forward to deliver a more natural and convenient AI voice experience."

This article was published with the assistance of generative AI and edited by The Korea Times.