{"slug": "voice-emotion-control-moves-from-ssml-to-prompts", "title": "Voice Emotion Control Moves From SSML to Prompts", "summary": "Kakao's Kanana-o model now accepts plain-language voice instructions such as 'read it in a sad voice' or 'read it in a Gyeongsang dialect', reflecting them in speed, volume, pitch, emotion, intonation, and intensity. On the Korean InstructTTSEval benchmark, Kanana-o scored 94.50, ahead of OpenAI's GPT-4o-mini-tts at 91.10 and just behind Google's Gemini 2.5 Flash Preview TTS at 95.38. The update also introduces LM-SPT, a tokenizer that compresses speech into fewer tokens for faster generation and lower serving cost.", "body_md": "Voice emotion control is leaving the markup. Kakao's Kanana-o model, detailed on its tech blog on August 4, 2026, now takes plain-language instructions like \"read it in a sad voice\" or \"read it in a Gyeongsang dialect\" and reflects them in speed, volume, pitch, emotion, intonation, and intensity. On the Korean InstructTTSEval benchmark it scores 94.50, ahead of OpenAI's GPT-4o-mini-tts at 91.10 and just behind Google's Gemini 2.5 Flash Preview TTS at 95.38.\n\nI work at Speechify, on the SpeechifyAI API side, so a post arguing that the markup layer is thinning is mildly inconvenient for my employer's docs team. Read it with that in mind. I think the direction is real anyway, and the trade-offs are worth thinking about before you build on it.\n\nKanana-o is Kakao's in-house omni model, developed by its Unified Foundation Model team. The August 4 update is about the speech layer: the model reads text naturally, and it now also follows delivery instructions written as ordinary sentences. \"Read it very quickly.\" \"Read it in a low voice.\" \"Read it in a Gyeongsang dialect.\" Kakao says the generated speech reflects speed, volume, and pitch, plus emotion, intonation, and intensity.\n\nThe interesting extension is role-based prompts. You can ask for a read \"like a sports broadcast\", \"like a news anchor\", or \"like reading a children's story\", and the model shapes the delivery to the scenario. It also stacks conditions in one instruction, such as \"lower the tone and read it quickly in a sad voice\". And despite training primarily on Korean data, Kakao says the same instructions work in English without difficulty.\n\nInstructTTSEval is a benchmark built to measure exactly this: how well a TTS model follows natural-language style instructions. The original, published by researchers in 2025, runs three tasks. Acoustic-parameter specification maps fine-grained cues like pitch and speed to audio. Descriptive-style directives take free-form style descriptions. Role-play gives the model a scenario and expects the delivery to match. Each task ships 1,000 English and 1,000 Chinese examples, judged automatically with an LLM-as-a-judge setup.\n\nKakao reports a Korean-language version of the benchmark, and its numbers are the concrete part of this announcement. Kanana-o scored 94.50. GPT-4o-mini-tts scored 91.10. Gemini 2.5 Flash Preview TTS scored 95.38. That puts Kakao ahead of OpenAI's model on instruction following and roughly level with Google's. It is also Kakao's own evaluation, run on Kakao's own setup, reported in Kakao's own post, so treat the exact gap as directional rather than gospel. The direction itself is the news: a messaging company's in-house model is competing on instruction-following benchmarks with the frontier labs.\n\nThe other half of the update is efficiency. Kakao built a tokenizer called LM-SPT, short for LM-aligned speech tokenizer, which compresses speech into fewer tokens before the language model processes it. Fewer tokens means less data to attend over, which means faster generation and lower serving cost for the same audio.\n\nKakao says LM-SPT outperformed models using Mimi, DualCodec, and CosyVoice2, three of the newer speech codecs, in its evaluations of Korean and English speech understanding and generation, and scored highest in expert listening tests on naturalness and speaker similarity. No public benchmark table for this yet, so that claim stands on Kakao's testing. The pattern is consistent with models absorbing the layers around them, which is the same fusion trend I wrote about in the Grok Voice analysis.\n\nHere is the part I actually care about. For the last decade, if you wanted a TTS voice to sound sad, you wrapped the text in markup. Speechify's API gives you 13 emotion presets through an SSML style tag, from warm to assertive to terrified, plus prosody, pauses, and pronunciation aliases, all in one request. That approach is deterministic. The same document produces the same delivery, you can version it, review it, and assert on it in tests.\n\nNatural-language instruction is the opposite trade. It is expressive, zero-markup, and readable by humans who will never touch SSML. It is also non-deterministic. Prompt phrasing changes output, models change under you, and you cannot diff an emotion the way you diff a tag. If your audio is a product, that is a real cost. The first teams to ship this will be the ones who treat the instruction as a testable artifact, the same lesson as pinning model versions, which I wrote about when Grok Voice repriced everyone through a floating alias.\n\nThe honest version also admits the direction of travel. Kakao's roadmap includes generating laughter, sighs, and exclamations, and folding speech understanding and generation into one architecture. Google and OpenAI are in the same race on InstructTTSEval. And the reason regional accents matter commercially is already measured: I wrote about the study where a regionally matched synthetic voice tripled brand recommendation over a neutral human read.\n\nSSML is not going away. Precise, scriptable, deterministic control has a long tail in production. But the default control surface for voice is shifting from markup to language, and the two will sit side by side for a while. Builders who keep their delivery instructions as data, rather than hardcoding either format, get to ride both.\n\nKanana-o is Kakao's in-house omni AI model, developed by its Unified Foundation Model team. On August 4, 2026, Kakao detailed an upgrade to its voice generation: the model follows natural-language delivery instructions like \"read it in a sad voice\" or \"read it in a Gyeongsang dialect\", plus role-based prompts and combined conditions. Kakao plans to apply it across its services. No public API has been announced.\n\nInstructTTSEval measures how well TTS models follow natural-language style instructions. It has three tasks: acoustic-parameter specification, descriptive-style directives, and role-play, with 1,000 English and 1,000 Chinese examples each, judged automatically. Kakao reported running a Korean version and scored 94.50, ahead of GPT-4o-mini-tts at 91.10 and close to Gemini 2.5 Flash Preview TTS at 95.38.\n\nSSML is deterministic markup: emotion tags, prosody, pauses, and pronunciation aliases produce repeatable output you can version and test. Natural-language control takes plain instructions and is more expressive, but output varies with phrasing and model updates. Speechify's TTS API exposes 13 SSML emotion presets, which suits production audio where the delivery needs to be pinned down.\n\nKakao's Kanana-o does, per its August 4, 2026 announcement. The InstructTTSEval benchmark also evaluates Gemini 2.5 Flash Preview TTS and GPT-4o-mini-tts, so Google and OpenAI are shipping instruction-following TTS too. For deterministic emotion control in production today, SSML-based APIs like Speechify remain the standard. The market is moving toward both.\n\nNot as of August 4, 2026. Kakao described the upgrade on its tech blog and said it plans to apply Kanana-o to its services going forward. No developer API, pricing, or availability date was announced. The benchmark numbers matter as a signal for the market, not as a product you can integrate today.", "url": "https://wpnews.pro/news/voice-emotion-control-moves-from-ssml-to-prompts", "canonical_source": "https://dev.to/lukeocodes/voice-emotion-control-moves-from-ssml-to-prompts-506b", "published_at": "2026-08-04 11:45:46+00:00", "updated_at": "2026-08-04 11:49:05.378963+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "natural-language-processing", "ai-research"], "entities": ["Kakao", "Kanana-o", "OpenAI", "GPT-4o-mini-tts", "Google", "Gemini 2.5 Flash Preview TTS", "Speechify", "InstructTTSEval"], "alternates": {"html": "https://wpnews.pro/news/voice-emotion-control-moves-from-ssml-to-prompts", "markdown": "https://wpnews.pro/news/voice-emotion-control-moves-from-ssml-to-prompts.md", "text": "https://wpnews.pro/news/voice-emotion-control-moves-from-ssml-to-prompts.txt", "jsonld": "https://wpnews.pro/news/voice-emotion-control-moves-from-ssml-to-prompts.jsonld"}}