Voice & Conversational Interfaces

Voice is invisible, linear, and ephemeral: users can't scan, can't see options, and can't re-read. Everything must be carried by dialog design. (For chat/LLM product UX beyond voice, see Designing AI-Powered Interfaces.)

Try it — reprompt escalation, not repetition. Reply with anything that isn't a weekday and watch the agent escalate: a brief hint, then examples, then a different channel entirely. The prompt is never repeated verbatim, and the copy never blames you.

Foundational theory

Grice's cooperative principle & maxims (1975) — the backbone of conversation design (Google's guidelines are explicitly built on it): be truthful (quality), say as much as needed and no more (quantity), be relevant (relevance), be clear/brief/orderly (manner). Most bad VUI copy violates quantity (over-long prompts) or manner (ambiguity).
Spoken language ≠ written language: contractions, shorter sentences, discourse markers ("Alright," "Got it") — write for the ear, read every prompt aloud (Pearl 2016; Google conversation-design guidance).

Dialog mechanics

Turn-taking: end prompts with a clear question or handoff; don't stack multiple questions in one turn.
Menus by voice: offer at most ~3 options per turn (working-memory limit is brutal without visuals); put the most likely/recommended option LAST (recency effect — the last thing heard is easiest to recall and repeat).
Confirmation strategy by stakes: implicit confirmation for low-risk ("OK, adding milk to your list"), explicit for high-risk ("Transfer $500 to Alex — should I go ahead?"). Never explicit-confirm everything; it destroys pacing (Cohen/Giangola/Balogh; Pearl).
Error handling — rapid reprompt escalation: first failure = brief reprompt with a hint; second = more guidance/examples; third = offer an alternative channel (screen, human, different task). Never repeat the identical prompt verbatim; never blame the user ("I didn't get that" not "You said that wrong").
Context & anaphora: retain conversational state ("What about tomorrow?" after a weather query must work); support corrections ("no, I meant Austin Texas").
Discoverability: the blank-slate problem — teach capabilities in context ("By the way, you can also…" sparingly), and answer "what can you do" with the top handful of useful actions, not a catalog.

The LLM era (2024–2026): what changed, what didn't

The classic assistants are being rebuilt on LLMs: Amazon announced Alexa+ (Feb 2025) with agentic, LLM-based skills and new AI-native SDKs; Google is replacing Assistant with Gemini on phones (transition announced 2025, completing into 2026); ChatGPT-style voice modes normalized fluid, open-ended spoken dialog.

Changed: intent grammars and rigid slot-filling give way to free-form understanding — users can phrase anything, interrupt (barge-in is now table stakes), and steer mid-answer. Streaming TTS lets you mask model latency by starting to speak early; acknowledge fast, then continue.
Not changed: working memory. Generated speech tempts systems into violating Grice's quantity maxim at scale — LLM verbosity is the new over-long prompt. Constrain spoken answers hard; push detail to a screen when one exists.
Raised stakes: models can misinterpret confidently or hallucinate, so confirmation-by-stakes matters more for real-world actions (money, messages, purchases, home control) — explicit confirmation before any consequential agentic act, with a clear readback of what will happen.
Persona: easier to break with generated text; enforce voice/style via system-level style guides and test transcripts, not hope.
Discoverability remains unsolved: "you can ask me anything" teaches nothing — still show/tell concrete, high-value examples in context.

Trust, transparency, privacy

Disclose non-human status; signal listening state clearly (audio/visual cue — a hardware norm across Alexa/Google/Siri); make mic controls and privacy behavior obvious; degrade gracefully to visual UI when available (multimodal: voice input + screen output is often the best pattern — "voice-forward," not voice-only, per Amazon's multimodal guidance).

Personas & speech quality

Define a consistent persona (word choice, prosody, humor level) matched to brand and task gravity; use SSML for pacing, emphasis, and pauses; latency: respond (or at least acknowledge with an earcon) fast — silence reads as failure.

Sources

Grice, H. P. (1975). "Logic and Conversation." In Syntax and Semantics 3. Academic Press.
Google — Conversation Design guidelines (developers.google.com/assistant/conversation-design). Note: the Conversational Actions platform these accompanied was sunset in June 2023, and Google Assistant itself is being replaced by Gemini on mobile (2025–2026); the design guidance remains sound and online but is effectively archival.
Amazon — Alexa Design Guide / voice & multimodal design documentation (developer.amazon.com/alexa); Amazon developer blog (Feb 2025) — Alexa+ and AI-native SDK announcement.
Pearl, C. (2016). Designing Voice User Interfaces. O'Reilly.
Cohen, M., Giangola, J. & Balogh, J. (2004). Voice User Interface Design. Addison-Wesley.
Nielsen Norman Group — intelligent-assistant and AI-chat usability studies (nngroup.com).