Voice Interaction Design (VUI)

Spoken interfaces have no screen to fall back on. The caller cannot scan a list, scroll up or see what they typed. Everything has to work in short bursts of audio with the caller holding one thing in memory at a time. VUI design is the set of decisions, from the greeting to the goodbye, that make a voice agent usable under those constraints. It sits above the models: the prompt, the flows and the error handling are all design artifacts.

How it works. The designer starts from the tasks callers actually have (book, reschedule, ask, report) and writes the ideal path for each as a short exchange, usually one question per turn. For every question they define the expected answers, the read-back format, and what happens on silence, on "I don't know", on an interruption and on a request to speak to a person. Those decisions become the system prompt, the tool definitions and the test scripts. The output is a conversation specification the engineering team can implement and the operations team can audit.

Why it matters for completion rate. Most voice agent failures are design failures, not model failures. A greeting that lists six options loses the caller by option four. A question that asks for three things at once gets one answer. A phone number read back as one long string cannot be checked. Good design raises task completion and lowers transfers without touching the model. It also sets expectations: callers who are told what the agent can do stay inside it.

Example. A gym's voice agent originally opened with "Welcome to Peak Fitness, I can help with memberships, class bookings, personal training, cancellations, opening hours and more. What would you like?" Callers froze or interrupted. The redesign opens with "Thanks for calling Peak Fitness. How can I help?" and lets the intent come from the caller. Class booking became one question per turn: which class, which day, which time, then a read-back. Numbers are confirmed in groups of three or four digits. Silence after a question gets one rephrase, then an offer to text a link.

Voice interaction design vs conversation design. Conversation design covers every channel, including chat, where the user can see history and tap buttons. Voice interaction design is the subset that deals with audio-only constraints: turn-taking, barge-in, pacing, pronunciation, and the fact that the caller cannot re-read anything. A chat flow ported to voice without that pass usually fails at the first long list.

Common pitfalls.

• Menus read aloud; speak one question and listen.
• Confirming everything; confirm only what triggers an action.
• Fillers on every turn, which slow the call and sound evasive.
• No path to a person; every flow needs an exit and a transfer rule.
• Testing with the designer's own voice only; test with real callers, accents and mobile audio.

Botcadence voice agents support natural barge-in and transfer to a person with the full transcript, which is where most of the design work above lands; see voice agents and the AI to human handoff design guide.