# Kakao Adds Prompt-Controlled Voice Styles to Kanana-o > Source: > Published: 2026-08-04 05:52:04+00:00 # Kakao Adds Prompt-Controlled Voice Styles to Kanana-o Kakao detailed improvements to Kanana-o on August 4 that let users steer speaking speed, volume, pitch, emotion and dialect with natural-language instructions. In the company's Korean InstructTTSEval test, the final checkpoint scored 94.50 after online reinforcement learning, up from 73.90 initially. The results are company-reported and do not establish independent real-world superiority. Kakao detailed improvements to its Kanana-o speech model on August 4, 2026, focused on letting users control how generated speech sounds with ordinary language. The company says prompts can adjust speed, volume, pitch, emotion, intonation and regional dialect, or request delivery styles such as a news anchor or sports broadcast. ### Controlling delivery, not only words The update combines Kakao's LM-SPT speech tokenizer with additional training for instruction-following. LM-SPT represents speech at 12.5 frames per second, half the frame rate of Kakao's previous tokenizer, while separating semantic tokens for spoken content from acoustic tokens for characteristics such as timbre and intonation. Kakao says this reduces the sequence length the model must generate and supports more detailed control over speech output. The company also describes a multi-codebook generation method that produces semantic and acoustic tokens in parallel before a decoder reconstructs the waveform. In Kakao's tests, the LM-SPT decoder used 40 million parameters and produced audio about 26 times faster than CosyVoice2 under the stated experimental conditions. Those figures come from Kakao's own evaluation and have not been independently reproduced in the retrieved sources. ### What the benchmark shows Kakao built a Korean benchmark using the InstructTTSEval framework to judge whether generated speech followed complex style instructions. The company reports that Kanana-o improved from 73.90 at its initial checkpoint to 81.30 after direct preference optimization and 94.50 after multi-objective online reinforcement learning. Kakao's table lists 91.10 for GPT-4o-mini-tts and 95.38 for Gemini-2.5-flash-preview-tts, making the final Kanana-o result competitive within this company-run test. The practical significance is controllability: developers building voice agents need predictable delivery as well as natural-sounding speech. But the post does not provide an independent evaluation, a public release schedule for the upgraded model or evidence from production use. Kakao says it plans to extend the work to nonverbal sounds such as laughter, sighs and exclamations and to apply Kanana-o across future services. ## Key Points - 1Natural-language prompts can control speed, volume, pitch, emotion, intonation, dialect and role-based speaking styles. - 2Kakao reports a Korean InstructTTSEval score of 94.50 after online reinforcement learning, compared with 73.90 at the initial checkpoint. - 3LM-SPT uses 12.5 speech frames per second and a smaller decoder, but the performance and efficiency results remain company-reported. ## Scoring Rationale The update provides concrete speech-model architecture and instruction-following results relevant to voice-agent developers. Its impact is moderated by company-run benchmarks, no independent reproduction and no confirmed public deployment schedule. ## Sources Primary source and supporting public references used for this report. Practice interview problems based on real data 1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with. [Try 250 free problems](/problems)