Natural-language commands now steer speech style and delivery

Custom voice tokenizer boosts generation speed and efficiency

Korean-language performance rivals global models; service rollout planned

Benchmark figures for Kakao's omni AI model Kanana-O [Provided by Kakao]
Benchmark figures for Kakao's omni AI model Kanana-O [Provided by Kakao]

Kakao has advanced the voice generation capabilities of Kanana-O, its in-house omni AI model. Users can now instruct the AI in natural language to adjust speech style, emotion and intonation, and the model reflects those preferences in detail when producing audio output.

Kakao said Tuesday it had upgraded Kanana-O's voice generation technology to let users control speech style, emotion and intonation through natural-language commands alone. The company also applied a proprietary voice tokenizer to improve generation speed and efficiency.

Kakao highlighted the model's ability to control vocal expression based on user instructions — a departure from conventional voice AI systems that focus primarily on reading text naturally. If a user says "read it very fast," "read it in a low voice" or "read it in a sad voice," the AI adjusts not only speed, volume and pitch but also emotion, intonation and intensity.

The model also responds to role-based instructions such as "like a sports commentator," "like a news anchor" or "like someone reading a picture book," as well as compound commands combining multiple conditions at once — for example, "read it fast in a low, sad voice."

In addition, Kakao said the model handles English voice generation with the same speech instructions without difficulty, despite having been trained primarily on Korean-language data.

Kanana-O scored 94.5 on the InstructTTSEval Korean benchmark, which evaluates a model's ability to follow speech instructions. That surpasses OpenAI's GPT-4o-mini-tts, which scored 91.1, and is comparable to Google's Gemini-2.5-flash-preview-tts, which scored 95.38.

Voice generation speed and efficiency have also improved. Kakao said it achieved this by applying LM-SPT, a proprietary voice tokenizer developed in-house. LM-SPT compresses audio into fewer tokens, reducing the volume of data the AI must process and enabling faster speech generation.

According to Kakao, LM-SPT outperformed models using the latest global technologies — including Mimi, DualCodec and CosyVoice2 — in evaluations comparing Korean and English speech comprehension and generation across multiple voice language models. It also received high marks in expert listening evaluations assessing the naturalness of synthesized speech and speaker similarity.

Kakao said it plans to keep developing Kanana-O's voice technology. The company will focus on advancing speech processing through research that handles both voice comprehension and generation within a single unified architecture, with the goal of delivering a natural and seamless voice experience. It also plans to develop the ability to generate nonverbal expressions such as laughter, sighs and exclamations.

"This upgrade to Kanana-O's voice technology focused on equipping the model not only to generate natural-sounding speech like a human, but also to express the tone, emotion and intonation a user wants based on natural-language instructions," said Noh Byeong-seok, performance leader for the Unified Foundation Model team at Kakao. "We will apply the Kanana-O model across a range of services to deliver a convenient AI voice experience."


chami@heraldcorp.com