콘텐츠로 이동

title: Conversational UI fundamentals (turn-taking, intents, context, modalities) applies_to: [conversational-ui, voice, chat, ai-chat] version: 1.0.0 last_updated: 2026-05 stability: stable

Conversational UI fundamentals

Conversational UI is design for back-and-forth exchanges. Voice assistants (Siri, Bixby), chatbots (customer support), AI chat (ChatGPT, Claude), and voice-controlled features in apps all share a foundation: a turn-taking exchange where each side has limited information about the other.

This file is the foundation for voice-ui-patterns.md, chatbot-design.md, ai-chat-interfaces.md, and korean-voice-conventions.md.

Modalities

Modality Examples
Voice-only Smart speakers (Echo, Nest), in-car assistants
Voice + visual Phone assistants (Siri, Bixby) — voice input, visual response
Text-only chat SMS bots, Slack bots, in-app chatbots
Multimodal AI chat ChatGPT (text + image upload + voice), Claude (text + image)
Hybrid (voice → text) Voice search → text results

Each modality has different constraints. Voice can't show a list of 20 options; chat can't convey vocal warmth. Pick the modality that matches the task.

Turn-taking

Conversation is turn-based:

User:    [request]
System:  [response]
User:    [follow-up or new request]
System:  [response]

The system must: 1. Detect end of turn — when has the user finished speaking / typing? 2. Respond appropriately — relevant, paced, complete. 3. Yield turn — signal "your turn" via prompt, cursor, or pause.

For voice: end-of-utterance detection is hard. Pauses can be thinking, not finishing.

For chat: end of turn is "Send" button or Enter. Easy.

Intents and slots (the canonical model)

Most voice / chatbot systems classify input via intents:

User: "Set a timer for 5 minutes"

Intent:        SET_TIMER
Slots:         duration: 5 minutes

The system maps utterance → intent + parameters → action.

Concept Definition
Intent What the user wants (action / question type)
Slot / entity Specific values (time, amount, name)
Context Prior turns; user state; system state
Confidence How sure the system is about the interpretation

For LLM-based systems (ChatGPT, Claude): intents are implicit. The model interprets natural language end-to-end. But for product features (smart-home control, customer support), explicit intent + slot architecture is cheaper / more reliable.

Conversational design principles

1. One thing per turn

Don't ask for two pieces of info in one turn:

✗ "What's your name and email?" ✓ "What's your name?" → user answers → "And your email?"

For visual chat: more info OK per turn (UI can hold a list). Voice: stricter.

2. Confirmation for destructive / costly actions

User: "Delete all my photos"
System: "Are you sure? This will delete 1,247 photos and can't be undone."
User: "Yes"
System: [does it]

Confirmation friction prevents accidents.

3. Graceful failure

When the system doesn't understand:

Bad Good
"Error" "Sorry, I didn't catch that. Could you say it differently?"
Silence Visible "Thinking..." indicator
Repeat exactly the same prompt Try a re-phrase or offer alternatives

After 2-3 failures: hand off to alternative (text input, human agent, settings menu).

4. Context-awareness

Remember context within a session:

User: "What's the weather in Seoul?"
System: "Sunny, 12°C."
User: "What about tomorrow?"   ← "tomorrow's weather in Seoul"
System: "Tomorrow in Seoul: cloudy, 9°C."

Don't make the user repeat "Seoul" every turn.

5. Be honest about capabilities

Don't pretend to know more than you do:

✗ "Yes, I can help you book a hotel" (when you can't) ✓ "I can search hotels, but you'll need to book on the hotel's site."

Misrepresentation breaks trust permanently.

6. Respect user time

  • Voice: don't make users sit through long preamble.
  • Chat: don't message-bomb (5 short messages = noise; one combined message is better).
  • AI chat: stream tokens so user sees response forming, not just a blank screen.

Streaming responses

For LLM-based chat, stream the response token-by-token rather than waiting for full completion:

[blank screen for 8 seconds] vs [text appears word-by-word starting at 200ms]

Streaming feels 10× faster even at the same total latency. Users see progress; mind fills in.

For voice: stream audio similarly (start speaking before full response is generated, hide TTS latency).

Latency budgets

Modality Acceptable latency
Voice (assistant) < 1 second to first audio
Chat (typing indicator) < 500ms to typing indicator
Chat (first response token) < 1 second
AI chat full response streamed; total 5-30s OK
Wake word detection < 200ms

Beyond the budget: users assume broken / abandon.

Affordances

For voice: how does the user know they can talk? - Wake word ("Hey Siri", "OK Google", "헤이 빅스비"). - Tap-to-talk button (visible mic icon, button press, holding). - Always listening mode (controversial; privacy concerns).

For chat: - Persistent input field at bottom. - Suggested prompts (chips above input) lower friction. - "Type a message..." placeholder with hint of capabilities.

Personality and voice

Conversational UI has personality. Define:

Dimension Examples
Formality Casual ("Hey!") vs Formal ("How may I assist you?")
Verbosity Terse vs Discursive
Warmth Professional vs Warm
Humor None vs Dry vs Playful
Emoji Never vs Rare vs Frequent

For Korean fintech: 해요체 friendly; minimal humor; emoji rare. For Korean entertainment: 해체 / 해요체 casual; some humor; more emoji. For Korean B2B: 합쇼체 formal; no humor; no emoji.

Pick deliberately and document. Don't let each engineer / writer pick differently.

Speech-to-text (STT) and text-to-speech (TTS)

For voice UIs:

STT (recognition)

  • Confidence scores — each transcription has a confidence; below threshold = "Sorry, I didn't catch that".
  • Language model bias — train STT on domain (medical, legal, fintech terms).
  • Noise robustness — test in real environments (kitchen, car, street).
  • Latency — STT runs in background; final transcript at end of utterance.

TTS (synthesis)

  • Voice choice — male / female / neutral; per-language.
  • Persona consistency — use one voice per assistant; users build relationship.
  • Prosody — emphasis, pauses, intonation matter. SSML allows control.
  • Latency — pre-render common phrases; stream long ones.

For Korean: Naver Clova, Samsung Bixby, Google TTS Korean, ElevenLabs Korean — all available. Test for naturalness; some still sound robotic.

Errors and recovery

Voice errors

Error Recovery
Couldn't hear "Sorry, could you say that again?"
Heard but didn't understand "I'm not sure what you meant. Did you mean X or Y?"
Understood but can't do "I can't do X. I can help with Y or Z."
System failure "Something went wrong. Try again later or say 'help' for options."

Chat errors

Error Recovery
Network error Retry button + offline indicator
AI response failed "Couldn't generate response. Try again?"
Rate limited "You've sent a lot of messages. Wait a minute."
Inappropriate content "I can't help with that. Try rephrasing."

After repeated failures: offer alternative (text input, human agent, web link).

Voice especially has privacy implications:

  • Always-on listening disclosure required.
  • Recording retention disclosure (where stored, how long, who can access).
  • Opt-in for voice training (using user audio to improve models).
  • Wake word false-positive transparency — Apple, Amazon publish how often Siri / Alexa false-trigger.

For Korean market: - Personal Information Protection Act (개인정보보호법) compliance. - Consent flows for voice recording / storage. - Right-to-delete for voice data.

For chat: - Disclose if conversations are stored / used for training. - Allow deletion. - Clear what's visible to humans (support agents, training reviewers).

Multilingual and code-switching

Korean conversational UI often has mixed Korean + English:

User: "맥북 프로 16인치 가격 알려줘"
User: "Schedule a meeting with 김철수 at 3pm"

System must: - Handle code-switching (mixed language input). - Preserve names / proper nouns in source language. - Choose response language (match user's primary, or per query).

Modern LLMs handle this natively. Older intent-based bots struggle.

Accessibility

Voice UI accessibility

  • Visual transcript of voice (deaf / hard-of-hearing).
  • Text alternative input for users who can't speak (motor disability, voice issues).
  • Visual feedback of audio (listening, processing, responding states).
  • Adjustable TTS speed.

Chat UI accessibility

  • Standard text accessibility (contrast, screen reader support).
  • Skip-to-content (long histories).
  • ARIA live regions for new messages.
  • Keyboard nav for messages (Up / Down to navigate).

For multimodal: provide both modalities. Voice-only excludes deaf users; text-only excludes low-vision / motor-impaired users.

Don't

  • Don't treat voice like a typed-command interface ("syntax errors" don't translate).
  • Don't ignore confidence scores — low-confidence guesses are worse than asking.
  • Don't bury error recovery in obscure phrases.
  • Don't break context within a session.
  • Don't lie about capabilities.
  • Don't force users into long preamble before action.
  • Don't leak privacy ("Did you say X?" loud in public is awkward).
  • Don't ship without graceful degradation when backend is down.

Cross-reference