콘텐츠로 이동

title: Voice UI patterns (smart speakers, in-app voice, IVR) applies_to: [voice-ui, smart-speaker, ivr, voice-assistant] version: 1.0.0 last_updated: 2026-05 stability: stable

Voice UI patterns

Voice user interfaces (VUIs) — smart speakers (Echo, Nest, NUGU), phone assistants (Siri, Google Assistant, Bixby, Clova), in-app voice features, IVR (interactive voice response phone systems), in-car voice control.

Read conversational-ui-fundamentals.md first.

Voice-only vs voice + visual

Modality Constraints
Voice-only (smart speaker, in-car) No visual; no list of options; rely on memory
Voice + visual (phone assistant) Voice input, visual response; richer feedback
Voice + ambient screen (Echo Show) Hybrid; visual supports voice

Pick the lowest common denominator. If your skill must work on Echo (audio only), don't design assuming visual.

Five voice UI patterns

1. Wake word activation

User: "Hey Siri, set a timer for 5 minutes."

Wake words: "Hey Siri", "OK Google", "Alexa", "Hey Bixby", "Hi Clova", "Hey NUGU".

Design: - Auditory feedback when wake word detected (chime, light pulse). - Visual feedback for visual-capable devices (animation, listening indicator). - Timeout — wake state lasts 6-10s; if no speech, returns to idle. - False positive handling — if wake fires accidentally, user can say "nevermind" or just stay silent.

2. Tap-to-talk

[User taps mic icon]
[Holds while speaking]
[Releases when done]

Or push-and-release (continuous listen until silence).

Design: - Visual press feedback during listening. - Visualization of audio level (waveform or pulse). - Cancel option (slide off button or back press). - Auto-end on silence (1-2 seconds of quiet).

3. Conversational follow-up

After response, briefly remain listening for follow-up:

User: "What's the weather?"
System: "62°F, sunny."
[brief pause; mic stays open]
User: "What about tomorrow?"
System: "Tomorrow: 58°F, cloudy."

Reduces "Hey Siri" repetition. Risk: false-trigger from background noise.

4. Multi-turn dialogue

User: "Book a flight to Seoul."
System: "Sure. What dates?"
User: "Next Friday, returning Sunday."
System: "Departing from where?"
User: "San Francisco."
System: "Found 12 flights. Cheapest is..."

System tracks slot-filling progress; asks only what's missing.

For complex tasks (flight booking, meal planning): multi-turn is necessary. Keep turns short; confirm at end.

5. Voice search → visual results

User: "Find me Korean restaurants nearby."
[voice → text → search query]
[visual results displayed]

Voice as input modality only; results are visual. Common in mobile apps (search by voice, browse results).

Smart speaker UX (audio-only)

For Echo, Nest, etc.:

Information density

You can convey ~15-30 words before user attention drifts. Be concise:

✗ "I found 5 Korean restaurants nearby. The first is Han Yang at 123 Main Street, with a 4.5-star rating, $$$ price, open until 10pm. The second is Bibim, located at 456..."

✓ "I found 5 Korean spots. Top result: Han Yang, 4.5 stars, half a mile away. Want details on it?"

Lists

Length Pattern
1 result "I found one: X."
2-3 results Read all briefly: "I found three. First: X. Second: Y. Third: Z."
4-7 results Read 2-3, ask: "Want to hear more?"
8+ results Summarize: "I found 12. Want me to read them?"

Don't read 12-item lists by default.

Audio cues

  • Earcon (short tone) for wake / acknowledgment.
  • Different chime for error vs success.
  • Voice timbre consistent across all responses.

Proactive notifications

Smart speakers can announce: - Reminders ("Your package arrived"). - Notifications ("Email from John").

Risks: - Noisy household: announcements during conversation = rude. - Privacy: "Email from John about your medical results" announced loudly.

Design: - Quiet hours setting. - Notification style option (silent vs chime vs full announcement). - Privacy filter — sensitive content not announced.

Phone assistant UX (voice + visual)

Siri, Google Assistant, Bixby on phones:

Visual response anatomy

┌─────────────────────────┐
│  [User: "What's the     │
│   weather in Seoul?"]   │   ← transcribed user input
├─────────────────────────┤
│                         │
│  [Sunny icon]           │
│  Seoul: 12°C            │   ← visual response
│  Sunny                  │
│  H: 15°C  L: 5°C        │
│                         │
│  [More details ↓]       │
└─────────────────────────┘

Voice spoken simultaneously: "Sunny, 12 degrees in Seoul."

Action buttons

Voice can suggest actions; visual provides buttons:

"I found a cafe nearby. Want directions?"
[Get directions]  [Call]  [Save for later]

User can voice-respond ("Yes") or tap.

In-car voice UX

Voice in cars (Apple CarPlay, Android Auto, native):

Constraints

  • Eyes on road: minimal visual demand.
  • Hands on wheel: voice or steering-wheel button.
  • Noisy environment: STT robustness + clear TTS.
  • Distraction limits: regulated in some jurisdictions.

Patterns

  • Steering wheel button activation (push-to-talk, no wake word).
  • Audio-only response by default; visual for confirmation.
  • Single voice command per task: "Call Mom" not "Open contacts and find Mom".

IVR (Interactive Voice Response — phone systems)

"For sales, press 1. For support, press 2."

Largely deprecated for chatbots / live agents in 2024+, but still common.

Best practices

  • Skip option to a human at any time ("Press 0 for an operator").
  • Avoid deep menus (max 3-4 levels).
  • Repeat option at end of menu.
  • Caller ID-based routing where possible (don't ask for account number you already know).

For Korean banking IVR: typically Korean voice menus, multi-tier, often with separate path for VIP customers.

Korean voice assistants

Assistant By Notes
Bixby Samsung Default on Galaxy phones; Korean strong
Clova / 클로바 Naver Smart speakers (Friends, Wave); Korean services
NUGU SKT Smart speakers (NUGU Mini, Candle); SKT services
GiGA Genie / 기가지니 KT Smart speaker; KT services
Kakao i Kakao Smart speaker (Kakao Mini); Kakao services
Siri / Google Assistant / Alexa International Korean support of varying quality

Korean assistants integrate with Korean services (Naver search, Kakao Talk, KT Olleh, SKT services). Western assistants integrate with global services (Apple, Google, Amazon).

For Korean B2C apps: voice features should consider Bixby Capsule integration if Samsung-priority audience.

Voice + accessibility

Voice UIs help motor / vision accessibility but exclude others:

Disability Voice UI helpful?
Motor (typing hard) Yes — voice replaces keyboard
Vision (can't see screen) Yes — voice replaces visual
Deaf / hard-of-hearing No — must offer text alt
Speech (can't speak) No — must offer text alt
Cognitive (slow processing) Variable — depends on pacing options

Always offer text alternative input AND visual output. Voice can be primary but never exclusive.

TTS voice selection

For your assistant's voice:

Choice Use
Single voice Most consistent; users build relationship
Voice options (male / female / neutral) More inclusive; users pick
Voice cloning (custom brand voice) Distinctive but expensive + ethics
Multilingual same-voice Consistent across locales

For Korean assistants: typically single-voice default; Korean female warm tone common (Bixby default).

SSML for prosody control

Speech Synthesis Markup Language adjusts TTS:

<speak>
  Hello! <break time="500ms"/> Today is
  <emphasis level="strong">Tuesday</emphasis>.
  The temperature is <say-as interpret-as="cardinal">72</say-as> degrees.
</speak>

Useful for: - Pauses / breaks. - Emphasis on key words. - Numbers vs digits ("1.0" → "one point zero" vs "one dot zero"). - Dates: "2024-05-08" → "May 8, 2024" or "5월 8일".

Korean SSML same syntax; supports Korean prosody marks.

Voice UI common mistakes

  • Too verbose responses — users abandon.
  • Lists too long — voice can't show 12 results.
  • No fallback when STT fails repeatedly.
  • Wake word false positives disrupting other audio.
  • Privacy violations announcing sensitive content.
  • Inconsistent persona (one response formal, next casual).
  • Ignoring code-switching (mixed Korean / English) for Korean users.

Don't

  • Don't make voice the only way to do things. Always offer touch / text alt.
  • Don't read 10-item lists by default. Summarize and offer to read.
  • Don't ignore false positives — wake words are noisy.
  • Don't drone on. Concise wins.
  • Don't leak privacy via voice ("Email from doctor about test results").
  • Don't break the "Hey Siri" ↔ "follow-up" expectation. Be consistent.
  • Don't ship one voice for all locales. Localize per-language.
  • Don't use only audio for critical actions (confirm with visual + voice when device permits).

Cross-reference