KaKao

- Fine-grained control over speaking style, pace, emotion, and intonation through natural-language instructions alone

- Enhanced speech generation quality, speed, and efficiency through a proprietary speech tokenizer

- Performance on par with leading global models on a Korean language benchmark, with plans for application to diverse services

 

[August 4, 2026] Kakao (CEO Shina Chung) announced that it has further advanced the speech generation technology of its in-house omni AI model, Kanana-o. According to a post published on Kakao’s tech blog on August 4, the company has enhanced Kanana-o’s speech generation technology to enable users to naturally control speaking style, emotion, and intonation through natural-language instructions alone. Kakao has also improved generation speed and efficiency by applying its proprietary speech tokenizer.

 

While conventional voice AI has primarily focused on reading text aloud naturally, the enhanced Kanana-o model allows users to control how speech is expressed based on their instructions. For example, when users provide natural-language instructions such as “Read it very quickly,” “Read it in a low voice,” “Read it in a sad voice,” or “Read it with a Gyeongsang-do dialect,” the AI generates speech that finely reflects not only speed, volume, and pitch, but also emotion, intonation, and emphasis.

 

Kanana-o can also follow role-based instructions such as “like a sports commentator,” “like a news anchor,” or “as if reading a storybook,” as well as compound instructions that convey multiple conditions, such as “Read it quickly in a lower, sad voice.”  Despite being primarily trained on Korean language, Kanana-o can also follow the same instructions when generating speech in English.

 

Kanana-o scored 94.50 on InstructTTSEval, a Korean benchmark designed to evaluate instruction-following capabilities in speech generation. The score surpassed GPT-4o mini TTS, which scored 91.10, and was comparable to Google’s Gemini 2.5 Flash Preview TTS, which scored 95.38.

 

In addition to improving speech generation quality, Kakao has significantly enhanced generation speed and efficiency. To achieve this, Kakao newly applied its proprietary speech tokenizer, LM-SPT (LM-aligned SPeech Tokenizer). LM-SPT is a technology that compresses speech into fewer tokens, reducing the amount of data that AI models need to process and enabling faster, more efficient speech generation.

 

In a comparative evaluation of Korean and English speech comprehension and generation capabilities of various speech language models, LM-SPT demonstrated superior performance to models incorporating leading global technologies such as Mimi, DualCodec, and CosyVoice 2.0. LM-SPT also achieved the highest score in an expert evaluation that assessed the naturalness of synthesized speech and speaker similarity through listening tests.

 

Kakao plans to continue advancing Kanana-o’s speech technology. The company will focus on further enhancing speech processing technology through research into an integrated architecture that handles speech comprehension and generation within a single framework, while working to deliver a more natural and seamless voice experience. Kakao also intends to continue developing more sophisticated speech-control technologies, including the natural generation of non-verbal expressions such as laughter, sighs, and exclamations.

 

Byungseok Roh, Head and Executive Lead of Unified Foundation Model Studio at Kakao, stated, “The latest advancement of Kanana-o’s speech technology focused not only on enabling natural human-like speech generation, but also on enabling the model to express desired speaking styles, emotions, and intonations based on users’ natural-language instructions. Going forward, we will apply the Kanana-o model to a range of services to deliver more natural and convenient AI voice experiences.” (E.O.D.)

 

Reference: Kakao Tech Blog https://tech.kakao.com/posts/827