Voice Interaction Flow

A voice interaction flow maps a spoken dialogue plus the machinery around it: wake word → listening state → intent recognition → confirmation → spoken response → next turn (or earcon, timeout, reprompt). It is a conversation flow (Conversation Flow) with three constraints added: there may be no visual channel, speech is ephemeral (users can't re-read the last turn), and the context is often hands-free and eyes-busy — driving, cooking, field work. Design evidence lives in Voice & Conversational Interfaces; this file covers how to map the flow.

When to use

Designing a voice assistant skill, in-car voice task, or smart-speaker interaction.
Any hands-free or eyes-busy workflow (field checklists, kitchen, clinical, warehouse) where speech is the primary channel.
Auditing a voice feature where users get stuck, go silent, or bail to the phone.

What it answers

How does each turn begin and end — wake word, follow-up mode, timeout?
What does the system say, exactly, and how brief can it be?
Which actions confirm implicitly vs. explicitly?
What happens on silence, misrecognition, and repeated failure?

What it includes (the voice-specific layer)

Wake/invocation — wake word or push-to-talk; whether follow-up turns re-require it; a clear listening-state signal (light/earcon — a hardware norm across Alexa/Google/Siri, per Voice & Conversational Interfaces).
Intent recognition step — including the "recognized but wrong" branch: support "no, I meant…" corrections, not just no-match.
Confirmation strategy by stakes — implicit for low-risk ("OK, adding milk"), explicit readback before consequential actions ("Transfer $500 to Alex — should I go ahead?"). Mark each action in the flow with its confirmation level; never explicit-confirm everything (Voice & Conversational Interfaces; Cohen/Giangola/Balogh).
Response brevity — every system turn is copy, written for the ear and read aloud; ≤ ~3 options per turn, most likely option last (recency); push detail to a screen when one exists ("voice-forward, not voice-only" — Amazon's multimodal guidance).
Reprompt escalation — no-input and no-match ladders at every turn: hint → examples → alternative channel/human. Silence is an event in the flow, not an absence.
Earcons & timing — short non-speech sounds for listening, success, and failure; acknowledge fast (an earcon beats dead air — silence reads as failure).
End of session — an explicit wrap-up line, not an open mic.

How it differs from a general conversation flow

No scanning, no persistent state on screen: the flow must carry context in memory-sized pieces; summaries and readbacks replace visible confirmation screens.
Ephemerality: users can ask "what were the options again?" — plan a repeat/where-am-I intent at every step.
Interruption is physical: hands and eyes are elsewhere; turns must survive long pauses (the user is holding a drone, not a phone) — set generous no-input timeouts in task flows, and re-orient on resume ("We were on step 3, propellers.").
Barge-in is table stakes in the LLM era: users interrupt mid-answer and steer; the flow should treat interruption as input, not error.

Worked example: hands-free checklist stepper

Flow shape: a checklist stepper — one item per turn, binary answers, system tracks position; the classic pattern for eyes-busy procedures.

User:   "Start my drone preflight checklist."
System: "Starting. Is the battery fully charged?"     [step 1]
User:   "Yes."
System: "Confirm propellers are secure."              [step 2]
User:   "Confirmed."
System: "GPS lock acquired? Say yes, no, or wait."    [step 3]
User:   (silence — user is checking the controller)
System: (after generous timeout) "Take your time. Say 'ready'
         when you have a GPS answer."                 [no-input, kept
                                                       open, no nag]
User:   "Yes, locked."
System: "Preflight complete. 6 of 6 checks passed. Safe flight."

Branch — a check fails: "No." → "Noted: propellers not secure.
  Fix and say 'recheck', or say 'skip' to log it and continue."
Branch — "Repeat that" / "Where am I?" works at every step.
Branch — no-match ×3 → "I'll pause the checklist. Open the app to
  continue on screen, or say 'resume' anytime."   [channel handoff]

Note what the shape gives you: fixed sequence (safety order matters), one fact per turn, resumability, and a spoken summary at the end.

Common mistakes

Prompts written for the eye, not the ear — long, formal, unreadable aloud (violates Grice's quantity/manner; the top VUI failure).
More than ~3 options per turn, or the recommended option first.
Identical reprompt repeated verbatim on failure.
No listening-state signal, so users talk to a closed mic.
Confirmation screens' job given to nothing: consequential actions fire without readback.
Treating silence as error instead of a state (eyes-busy users pause).
Voice-only when a screen exists — no visual fallback or handoff.

Checklist

Wake/invocation and listening signal defined; follow-up mode set
Every system turn ≤ one question; copy read aloud and timed
≤ 3 options per turn; most likely last
Confirmation level (implicit/explicit) marked on every action
No-input AND no-match ladders at every turn; never verbatim repeat
Repeat / where-am-I / go-back intents work at every step
Barge-in and mid-answer corrections handled
Earcons for listening/success/failure; fast acknowledgment
Handoff to screen or human defined, with context transferred
Session end is explicit; resumable tasks re-orient on return

Sources

Google — Conversation Design guidelines (developers.google.com/assistant/conversation-design) — turn-taking, error ladders; archival but sound.
Amazon — Alexa Design Guide / voice & multimodal design documentation (developer.amazon.com/alexa) — confirmation, brevity, voice-forward multimodal.
Pearl, C. (2016). Designing Voice User Interfaces. O'Reilly.
Cohen, M., Giangola, J. & Balogh, J. (2004). Voice User Interface Design. Addison-Wesley — confirmation strategies, reprompt design.
Grice, H. P. (1975). "Logic and Conversation." Syntax and Semantics 3.
Nielsen Norman Group — intelligent-assistant usability studies (nngroup.com).