콘텐츠로 이동

title: Korean voice and conversational conventions applies_to: [korean, voice, chatbot, conversational, kr-market] version: 1.0.0 last_updated: 2026-05 stability: stable

Korean voice and conversational conventions

Korean voice and chat UIs have specific market patterns, regulatory requirements, and language conventions that differ from Western defaults.

Read conversational-ui-fundamentals.md first.

Korean voice assistant landscape

Product By Strength
Bixby Samsung Default on Galaxy phones; deep Korean language support; Bixby Capsules for 3rd-party integration
Clova / 클로바 Naver Smart speakers (Friends, Wave, Slate); HyperCLOVA-powered
NUGU SK Telecom Smart speakers (NUGU Mini, Candle); SKT services integration
GiGA Genie / 기가지니 KT Smart speakers; KT services
Kakao i Kakao Smart speaker (Kakao Mini); Kakao services (Talk, Map, T)
HyperCLOVA X Naver Korean-first LLM; conversational AI
Apple Siri Apple iOS; Korean support adequate
Google Assistant Google Android; Korean strong
Amazon Alexa Amazon Limited Korean market

For B2C apps targeting Korean users: prioritize Bixby (Samsung's market share in KR is large), Clova (Naver ecosystem reach), and Google Assistant (Android global).

Honorific level (존댓말)

The most important Korean conversational design choice:

Level Suffix Use
합쇼체 ~ㅂ니다 / ~습니다 Formal; banking, government, premium B2B
해요체 ~해요 / ~예요 Friendly polite; consumer apps, KakaoTalk channel bots
해체 ~해 / ~야 Casual; rare in commercial UI; kid / friend characters

Default for product type

Product Default
Banking app voice 합쇼체
Fintech consumer (Toss, KakaoBank) 해요체
E-commerce chatbot 해요체
Customer support (formal company) 합쇼체
Customer support (casual brand) 해요체
Kids' content 해요체 (slightly softer)
Government / civic 합쇼체
Smart speaker (Bixby, NUGU) 해요체
Voice navigation (T-map, Naver Map) 해요체

Don't mix levels

Switching between 해요체 and 합쇼체 in one bot's responses sounds inconsistent / sloppy. Pick one and apply throughout.

Exception: agent-handoff transitions where company voice (합쇼체) replaces bot voice (해요체) is acceptable as a deliberate handoff signal.

Mixed in long-form content

For longer responses (FAQ answers, terms), 합쇼체 sometimes fits even when surrounding chat is 해요체:

User: 환불은 어떻게 받나요?

Bot: (해요체 conversational layer)
환불 절차를 안내해 드릴게요!

(합쇼체 formal section, often quoted/styled differently)
"주문일로부터 7일 이내에 반품 신청을 하셔야 합니다.
반품 신청 후 7일 이내에 배송이 회수되며..."

(다시 해요체)
더 궁금한 점이 있으면 알려주세요!

Greeting patterns

Time of day Greeting
Anytime 안녕하세요!
Morning 좋은 아침이에요!
Late evening 늦은 시간 죄송해요 (acknowledge after-hours)

For chatbot greeting: - "안녕하세요! [브랜드] 챗봇이에요. 무엇을 도와드릴까요?" - "어떤 도움이 필요하세요?" - "[브랜드]에 오신 것을 환영해요!"

For voice assistant: short, no preamble — "네, 듣고 있어요" or just chime.

Confirmation phrases

Action Korean
Affirmation "네", "확인했어요", "알겠어요"
Confirmation request "...하시겠어요?" / "...할까요?"
Reassurance "잠시만 기다려 주세요" / "확인 중이에요"
Apology "죄송해요" / "죄송합니다" / "다시 말씀해 주시겠어요?"
Thanks "감사해요" / "감사합니다"

Apology when bot doesn't understand

"죄송해요, 잘 못 알아들었어요. 다시 말씀해 주시겠어요?"
"제가 잘 이해하지 못했어요. 다른 방식으로 말씀해 주시겠어요?"
"음, 잘 모르겠어요. 도움이 필요하시면 [메뉴/연락처] 확인해 주세요."

Code-switching (mixed Korean + English)

Korean speakers often code-switch:

"맥북 프로 16인치 가격 알려줘"        ← English brand + Korean
"Schedule a meeting with 김철수"      ← Korean name in English request
"오늘 저녁에 LA에 도착해"             ← English city in Korean

Bot must: - Handle mixed-language input. - Preserve proper nouns in source language. - Choose response language (default: match user's primary).

LLM-based bots (HyperCLOVA, GPT-4, Claude) handle this natively. Older intent-based bots struggle.

Korean-specific speech patterns

반말 (informal) detection

If user uses 반말 ("말해줘"), the bot must NOT respond in matching 반말 — looks rude. Stay in 해요체 / 합쇼체 always.

If brand voice is intentionally casual (Toss-style "money character" voice), 해요체 is the friendly endpoint.

Honorifics for the user

Bot should use 님 to address the user: - "고객님 안녕하세요" - "[Username]님, 주문이 도착했어요!" - 회원님 / 사용자님 / 고객님 (varies by brand)

Not addressing the user

  • "주문이 도착했어요" (good — passive)
  • "당신의 주문이 도착했어요" (bad — 당신 is impersonal / formal-distant)

당신 is rarely used in Korean commercial UI. Avoid.

Voice and TTS in Korean

Voice characteristics

Most Korean assistants default to: - Female, gentle. - 20s-30s perceived age. - Clear enunciation. - 해요체 register.

Reasons: - Cultural expectation (legacy from Japanese / Korean GPS / station announcements). - Perceived warmth / approachability. - Test results (most users prefer female voice for assistants in studies).

Modern apps offer choice: male voices, neutral voices, celebrity voices (premium).

TTS quality

Service Korean TTS quality
Naver Clova Voice Strong; many voice options
Samsung Bixby TTS Strong; Galaxy default
Google Cloud TTS Good; multiple Korean voices
ElevenLabs Improving; voice cloning for Korean
Apple Siri TTS Adequate; iOS only
AWS Polly Adequate
Typecast Korean-native; AI voice for content creation

For brand voice: use Clova or Bixby for Korean-first; ElevenLabs for cloning.

Pronunciation challenges

Hangul → speech challenges: - Number reading: 1,000 = "천" or "1천" — context-dependent. - Date format: "5월 8일" not "⅝". - Time: "3시" not "3:00". - Currency: "5,000원" → "오천 원" (read), not "5,000 won". - English brand names: "Coca-Cola" → "코카콜라" or "코카-콜라" depending. - Acronyms: read letter-by-letter or as Korean (e.g., NASA → "나사").

Modern Korean TTS handles most of this; test edge cases.

STT (speech-to-text) Korean

Quality: - Naver Clova Speech — strong for Korean. - Bixby STT — strong; Galaxy ecosystem. - Google Cloud Speech — strong; supports code-switching. - OpenAI Whisper — open-source; excellent Korean. - Apple speech recognition — adequate.

Issues: - Dialect (Busan, Jeju) — quality varies; Whisper handles best. - Noise — Korean characters are dense; mishears more common in noisy environments. - Code-switching — depends on STT model. - Numbers in speech — context-dependent ("이천이십사 년" = 2024).

Korean regulatory

Personal Information Protection Act (개인정보보호법)

For voice / chat that collects personal data: - Consent required for collection. - Disclosure of what's collected, why, how long stored, with whom shared. - Right to delete voice recordings + chat history. - Encrypted storage for sensitive data.

Voice assistants must disclose recording in privacy policy. Some require explicit "always listening" opt-in.

Information network Act (정보통신망법)

For chatbot interactions: - Same rules as email marketing if chatbot proactively pushes offers. - 광고 disclosure if chatbot promotes paid content.

Financial product chatbot (자본시장법 / 금융소비자보호법)

If bot discusses investments / financial products: - Risk disclosure required. - Advice qualifications — bot can't give personalized investment advice without licensing. - Consumer protection disclosures.

Korean fintech bots typically include extensive disclaimer text.

KakaoTalk channel chatbot specifics

KakaoTalk is dominant in Korea; many businesses run chatbots there:

  • Use Kakao i Open Builder OR Kakao Channel admin tools.
  • 답변 시간 (response time hours) displayed.
  • Persistent menu (always-visible bottom bar) with categories.
  • Auto-greeting on first message.
  • Can hand off to human within KakaoTalk.

Design: - Stick to KakaoTalk message types (text, image, card, list). - Don't replicate native UI in card format — looks broken. - Test on KakaoTalk mobile app primarily (most users there).

Common Korean chatbot / voice mistakes

  • Mixed honorific levels — 해요 + 합쇼 in same conversation.
  • Wrong register for brand — bank using 해체 / casual register.
  • Direct translation from English — sounds unnatural.
  • 당신 used as "you" — too formal-distant.
  • No 님 attached to user — feels rude.
  • English brand names not localized — "Apple Pay" vs "애플 페이" inconsistency.
  • TTS mispronouncing user names — verify common names.
  • Long preamble — Korean voice users want directness.

Don't

  • Don't mix 해요체 and 합쇼체 within one bot's voice. Pick one.
  • Don't translate Western chatbot scripts word-for-word. Adapt to Korean phrasing.
  • Don't use 당신 as default "you". Use 고객님 / 회원님 / username + 님.
  • Don't ignore code-switching. Korean users mix English freely.
  • Don't ship voice in Korean without testing TTS pronunciation on brand / product names.
  • Don't omit 개인정보보호법 disclosures for data-collecting voice / chat.
  • Don't use 반말 in commercial UI even if user uses it. Stay polite.
  • Don't make Bixby Capsule integration an afterthought for Samsung-priority audience.

Cross-reference