Kakao Kanana-o and the Instruction Compliance Trap: Why Following Voice Prompts Is Not the Same as Correct Output

Kakao announced on August 4, 2026 that its omni-AI model Kanana-o now supports natural language control over voice tone, emotion, intonation, speed, volume, and pitch. Users type prompts like "read this quickly in a sad voice with a lower tone" or "read like a sports commentator," and the model adjusts its spoken output accordingly. Kanana-o scored 94.50 on the Korean category of the InstructTTSeval benchmark, beating OpenAI's GPT-4o-mini-tts at 91.10 and approaching Google's Gemini 2.5 Flash Preview TTS at 95.38. But instruction compliance and output correctness are two different guarantees, and production teams confuse them at their own risk.

What Does Kakao Kanana-o Actually Do?

Kanana-o is Kakao's proprietary AI model upgraded to generate speech that responds to natural language style instructions. According to the Korea Times, it handles single-attribute prompts ("read in a low voice"), role-based prompts ("read like a news anchor"), and multi-condition prompts that combine several attributes at once. The model also supports Korean regional dialects, including Gyeongsang-do and Jeolla-do, and executes equivalent instructions in English despite being trained primarily on Korean data.

Kakao integrated a new self-developed voice tokenizer called LM-SPT that compresses speech into fewer tokens, reducing processing load and increasing generation speed. The company plans to unify voice understanding and generation into a single architecture and develop non-verbal expression capabilities like laughter, sighs, and exclamations.

What Does InstructTTSeval Actually Measure?

InstructTTSeval, introduced in a June 2025 paper on arXiv, benchmarks whether a TTS model follows the style instructions it receives. It evaluates compliance with requested attributes: did the output sound sad when you asked for sad? Did it speed up when you asked for faster? Did it lower the pitch when prompted?

It does not measure whether the audio pronounced "Kakao" correctly. It does not check whether a phone number was read digit-by-digit or collapsed into a misleading phrase. It does not verify that the same prompt produces the same output twice. Instruction compliance is about style. Output correctness is about content.

A score of 94.50 means Kanana-o follows style instructions correctly 94.5% of the time on the benchmark's test set. That leaves 5.5% non-compliance on curated prompts. In production at 10,000 clips, that projects to 550 outputs that did not match the requested style. The number says nothing about how many of those clips also mispronounced a proper noun or misread a date.

Why Does Natural Language Voice Control Add Variance?

Every control surface you give a probabilistic model is a new dimension of variance. SSML tags like <prosody rate="fast"> are deterministic instructions: the model interprets "fast" through a fixed mapping. Natural language prompts like "read quickly in a sad voice with a lower tone" are ambiguous by design. "Quickly" has no fixed speed value. "Sad" has no fixed pitch contour. "Lower tone" interacts with "sad" in ways the model resolves probabilistically each time.

Deepgram's production testing identified five categories of TTS pronunciation errors that persist regardless of style controls: proper noun failures, number formatting errors, acronym misreads, homograph confusion, and foreign word mispronunciation. Adding emotion and dialect knobs does not reduce these errors. It layers additional variance on top of them.

Multi-condition prompts compound the problem. "Read like a sports commentator in a Gyeongsang dialect at high speed" combines role, dialect, and pace into a single instruction. The model resolves the interaction between these constraints differently on each generation. Two runs of the same prompt can produce measurably different outputs, and neither run guarantees the content is pronounced correctly.

What Happens When You Ship Instructed Voice at Scale?

Content teams adopting Kanana-o or similar instruction-following TTS models face a specific production failure: they validate the style once and assume it holds across thousands of clips. It does not.

The Princeton Generative Engine Optimization study found that adding verifiable statistics to content boosts AI visibility by up to 40%. The implication for voice production is parallel: verifiable, measurable quality attributes (pronunciation accuracy, format compliance, voice consistency) are what distinguish production-ready audio from demo-quality audio. Style compliance is not one of those attributes.

Consider a customer service IVR that uses Kanana-o to generate empathetic responses. The instruction "read in a warm, reassuring voice" might score well on InstructTTSeval. But if the model mispronounces the customer's bank name, reads a dollar amount as a date, or drifts voice identity between the greeting and the account summary, the empathetic tone becomes irrelevant. According to Sarvam AI's Bulbul V3 benchmarks, word skips and mispronunciations persist even in models that score highly on naturalness, because naturalness and accuracy are measured on different axes.

How Do You Close the Gap Between Compliance and Correctness?

The fix is not to avoid instruction-following TTS. It is to validate what instruction-following benchmarks do not measure. Four production requirements apply:

1. Lock a pronunciation reference per language and dialect. Kanana-o supports Gyeongsang-do and Jeolla-do dialects plus English. Each dialect is a separate failure surface for proper nouns, numbers, and domain vocabulary. A locked pronunciation dictionary per locale catches errors that style instructions cannot.

2. Pin the model version. Kakao plans to continue developing Kanana-o, including adding non-verbal expressions. Each update changes the model's behavior under the same prompt. Pin the version you validated. Test before migrating.

3. Score every output against a reference. A 94.50 InstructTTSeval score is a population average on a fixed test set. Production quality lives in the tail. Score every clip for pronunciation accuracy, voice identity match, and format compliance before it ships.

4. Regenerate only failures. When a clip fails validation, regenerate that specific clip. Do not re-run the entire batch. Targeted retry keeps cost proportional to actual failures, not batch size.

Onepin is a voice workflow platform that orchestrates, validates, and ships production-ready audio across 100+ TTS models. It applies these four checks to every output regardless of which model generated it, closing the gap that instruction compliance benchmarks leave open.

The Benchmark Proves the Model Listens. Production Requires Proving It Spoke Correctly.

Kakao Kanana-o represents genuine progress in making voice AI respond to human intent. Natural language voice control is more intuitive than SSML markup and more flexible than preset voice profiles. A 94.50 score on InstructTTSeval puts Kanana-o in competitive range with OpenAI and Google on instruction following.

But instruction following is the input side of the problem. Output correctness is the delivery side. The benchmark proves the model listens. Production requires proving it spoke correctly, every time, at scale, across every language and dialect it supports. That proof does not come from the model. It comes from the production layer above it.