Skip to main content Scroll Top

Voice User Interface Design: Building Voice UI for Apps, Assistants and Devices

Updated November 2026 · Written and maintained by the Progression Agency strategy team

Voice user interface design is the work of deciding what a spoken product says, what it listens for, and how it recovers when it mishears, so a person can finish a task without a screen. It covers the intents a voice app understands, the prompts it speaks, the confirmations it gives, the error paths it takes and how it hands off to a display when one is there. Amazon, Apple and Google each publish rules for voice interfaces on their platforms, and those rules shape the design before any code is written. Progression’s designers, developers and strategists are based in New York and build voice experiences for companies across the US.

On this page · 16 sections
  1. A booking line that read out nine time slots
  2. What voice user interface design decides
  3. One conversation, turn by turn
  4. Intents, utterances and slots: the vocabulary of a voice UI
  5. Prompts written for the ear
  6. When the system mishears: No Match, No Input and system errors
  7. Confirmations: implicit, explicit or none
  8. Voice with a screen: multimodal design
  9. Voice interfaces and accessibility
  10. Alexa skill development: what Amazon requires
  11. Siri and App Intents on Apple devices
  12. Google Assistant after Conversational Actions
  13. Personas and sample dialogs before any build
  14. How is a voice UI tested with real speech?
  15. What voice interface work costs
  16. How Progression designs and builds voice interfaces

Mostly typed, not spoken? For an assistant that mostly types rather than talks, our AI chatbot development page covers text chat built on your own content.

Examples marked “Illustrative example.” are hypothetical, and no real business, product or voice app is shown.

A booking line that read out nine time slots

Picture a hypothetical bike repair shop that adds voice booking to its phone line. The first version answers a request by reading every open time in one breath, nine of them, and then asks which one the caller wants.

Two drawn boxes. The first shows a hypothetical voice booking prompt reading nine times in one breath, which callers could not hold in memory. The second shows a narrowing question about morning or afternoon, then three choices and a read back.Two drawn boxes. The first shows a hypothetical voice booking prompt reading nine times in one breath, which callers could not hold in memory. The second shows a narrowing question about morning or afternoon, then three choices and a read back.
Illustrative example. A hypothetical voice prompt that listed nine times, and the version that asked a narrowing question first.

The recognition worked and the calendar worked. The prompt was the problem, because a listener cannot scroll back. The second version asks a narrowing question first, which is what Apple’s Siri guidelines recommend when a full list is too long to read in a reasonable time.

  • Count how many options a prompt reads aloud.
  • Narrow with one question before listing.
  • Read back the choice before acting on it.
  • Keep the same flow working when the caller interrupts.

What voice user interface design decides

Google’s conversation design guidance describes building for Assistant as a breadth of skills, voice user interface design among them, alongside interaction design, visual design, motion design and UX writing, refined into one discipline called conversation design. In practice the work splits into layers, each with its own failure when it is skipped.

The layers of a voice user interface and what each decides
LayerWhat it decidesWhat goes wrong without it
IntentsThe tasks the product can do when askedRequests land nowhere, or in the wrong task
Utterances and slotsThe many ways people phrase a request, and the details inside itPeople must guess the exact words
PromptsWhat the system says, in what order and how brieflyLong menus nobody can hold in memory
ConfirmationsWhich details are read back, and when to ask yes or noWrong orders, or tedious double checks
Error pathsWhat happens on silence, a miss or a failed systemPeople repeat themselves and give up
Screen handoffWhat also appears on a display, and whenA spoken reply that depends on a picture
Persona and toneA consistent voice across every promptPrompts that sound like different products

Is a voice user interface the same as a chatbot?

They share intents and dialog logic, but a voice user interface is heard rather than read, so it has to work without scrolling, links or a list on the screen. Prompts get shorter, choices get narrower, and every error path has to assume the person may not have heard the last thing said. Our AI chatbot development team handles the text side.

One conversation, turn by turn

The clearest way to see a voice UI is as a sequence of turns. Each turn has something the system says, a range of things a person might say back, a decision the design makes, and a plan for when it fails.

Illustrative example. A hypothetical voice booking, turn by turn
TurnWhat the system saysWhat the person may sayWhat the design decidesIf it fails
InvocationA short welcome naming the shop“Book a tune up”The invocation name and the first intentOffer help on what can be booked
Narrowing“Morning or afternoon?”“Afternoon”, “after lunch”, “around two”A time of day slot with synonymsRepeat with two examples
Choice“I have 1, 1:30 or 2. Which one?”“One thirty”, “the middle one”Slot values and ordinal phrasesRead the three times again, shorter
Confirmation“A tune up Thursday at 1:30. Shall I book it?”“Yes”, “No, Friday”Explicit confirmation, since a booking is hard to undoAccept a one step correction
Result“Booked. A text with the details is on its way.”“Thanks”Where the details are sent, and saying soName another way to book
ExitNothing more, the session closes“Stop”, “cancel”Stop and cancel handled in every stateClose cleanly
A six step figure: invocation, open prompt, utterance matched to an intent with slots, confirmation, fulfillment and exit, each with one line on what it decides.A six step figure: invocation, open prompt, utterance matched to an intent with slots, confirmation, fulfillment and exit, each with one line on what it decides.
Voice user interface design, drawn as the six turns of one spoken booking.

Writing the turns as a table before any build exposes the gaps: a turn with no failure plan, or a confirmation where none is needed. Our AI receptionist development work starts from the same kind of turn map for phone and web intake.

Intents, utterances and slots: the vocabulary of a voice UI

Amazon’s interaction model documentation gives the clearest names for the parts. You define the voice interface by mapping what people say to the intents your service can handle, and an intent can carry slots for the details inside a request. Amazon’s own example is a trip planner with slots for the starting city, the destination and the travel date.

The parts of an Alexa interaction model
Alexa interaction model partWhat Amazon’s documentation says it is
IntentAn action that fulfills a spoken request, defined in a JSON intent schema
SlotAn optional argument of an intent, such as a city or a date, with a built-in or custom type
Sample utterancesLikely spoken phrases mapped to each intent, with slots in curly brackets
Custom slot typeA representative list of values for items no built-in type covers
Dialog modelOptional steps for a multi turn conversation that collect, validate and confirm slot values

Built-in slot types do the heavy lifting for common details. Amazon’s date type, for example, turns words such as today or next Friday into a date, so the design can focus on the parts that are specific to the business.

How many sample utterances does an intent need?

As many representative phrases as you can gather, in Amazon’s words, and every one must lead to a relevant, grammatical response in the language of the locale. Start from how customers already phrase requests on calls, in chat logs and in search terms, then add the short and the rambling versions.

  • The plain request: book a tune up.
  • The detailed request: book a tune up for Thursday afternoon.
  • The indirect request: my brakes are squeaking.
  • The fragment: Thursday.
  • The correction: no, Friday.

Prompts written for the ear

Apple’s Siri Human Interface Guidelines read like a style guide for spoken prompts. Responses should be clear and as succinct as possible, because people hear them again and again; humor and extra words wear thin. A spoken reply should stand on its own without a visual element, and it should not depend on which device answered.

Seven prompt checks drawn from Apple's Siri Human Interface Guidelines.Seven prompt checks drawn from Apple's Siri Human Interface Guidelines.
A drawn card of seven prompt checks: specific, succinct, standing alone, device independent, no app name, inclusive and specific errors, with a closing line about narrowing long lists.
  • Ask a specific follow up: “Which soup?” is clearer than “Which one?”.
  • Leave out the app’s name; the system already attributes the reply.
  • Avoid guessing a pronoun: “Who should I send it to?”.
  • Keep advertising and in-app purchase pitches out of spoken content.
  • Replace a long list with an open question that narrows it.
  • Make error replies specific to what went wrong.

Why do short prompts beat complete ones?

Because a listener holds only a few items at once and cannot look back. A short prompt that leads to a second short prompt usually finishes faster than one long prompt that has to be repeated.

When the system mishears: No Match, No Input and system errors

Google’s guidance on errors sorts them into three kinds: No Match, when the response cannot be understood in context; No Input, when nothing was heard while the microphone was open; and system errors, when a service the app depends on cannot finish the task. Google puts the stakes plainly: one poorly handled error can outweigh dozens of successful interactions.

A four step figure for error handling: a quick reprompt on the first no match, examples on the second, a graceful ending after two, and a rephrased question when there is no input at all.A four step figure for error handling: a quick reprompt on the first no match, examples on the second, a graceful ending after two, and a rephrased question when there is no input at all.
An error path for a voice app, from the first miss to a graceful ending.
  • Assume the person is cooperating and what they said was relevant.
  • Never repeat the original prompt word for word.
  • Add options or examples on the second attempt.
  • End the conversation after two No Match attempts.
  • Point to another way to finish the task when the voice path cannot.
  • Be honest about why something did not work.

What should a voice app say when it hears silence?

It should ask again in a rephrased, concise way, written as if the person never heard the question. Google advises against saying “I didn’t hear you”, since the person may not have said anything, and suggests offering options when they may be unsure what to say. When the detail is optional, the dialog can simply move on.

Confirmations: implicit, explicit or none

Google’s confirmations guidance treats read backs as a tool with a cost. They let people correct mistakes at once and keep the thread of the conversation, but belabored confirmations slow a voice app down.

Four kinds of confirmation and when each fits
Kind of confirmationWhen Google’s guidance uses itExample in a booking
ImplicitMost of the time, to carry the details said or implied“For Thursday at 1:30…” and move on
ExplicitBefore an action that is hard to undo, or when a misunderstanding costs a lot“Shall I book it?” before booking
AcknowledgmentAfter a completed action that is not self evident“Booked, and the details are on their way by text”
NoneSimple, high confidence input, such as yes or noNo “OK, yes” after a yes

Expect corrections after any confirmation. People tend to answer with a no followed by the right value, such as “No, 7 AM”, which Google calls a one step correction, and the design should accept it without starting over.

Planning a voice feature?Tell us the task people should finish by voice and the devices they use. You get a free proposal with the intents and the first sample dialogs.

Get a free proposalTell us about a project you’d like help with.

Voice with a screen: multimodal design

Many voice interfaces now run on devices with a display, from smart speakers with screens to phones and cars. Apple’s guidelines ask for responses that Siri can deliver both audibly and visually, so the system can choose: a weather forecast appears on an iPhone screen and is spoken through AirPods. Amazon’s Alexa Design Guide lists being multimodal among its voice design principles, beside being natural, brief, contextual and trustworthy.

  • Say the essential fact; show the supporting detail.
  • Never make the spoken reply depend on the picture.
  • Let touch and voice reach the same result.
  • Keep on-screen labels in the words people would say aloud.

Screen layouts and touch flows for the same product sit with our UI/UX design services and mobile app design teams.

Voice interfaces and accessibility

Voice interfaces help many people, and they also have to respect people who speak to controls they can see. The W3C’s Understanding Success Criterion 2.5.3, Label in Name (Level A) asks that, for components labeled with text, the accessible name contains the visible text, because people who use speech recognition operate controls by saying the labels they see.

Do on-screen buttons need to match what people say?

They should. If a button reads Send but its accessible name is something else, a speech input user who says “click Send” fails. The W3C notes that putting the label’s text at the start of the name is best practice, and that mismatches add load for people who also have cognitive challenges.

  • Match every visible label to the control’s accessible name.
  • Offer a non voice path for every voice task.
  • Allow time to answer, since the reprompt window is short.
  • Write prompts without assumptions about gender or ability.
  • Test with people who use speech input every day.

Websites that need the same care get it from our ADA-compliant website design team, and research sessions with real users run through our usability testing services.

Alexa skill development: what Amazon requires

Alexa skill development follows rules that turn into design constraints. Amazon’s built-in intents page says a skill must implement AMAZON.StopIntent, with the session ending on it, while AMAZON.HelpIntent explains how to use the skill and AMAZON.FallbackIntent catches requests that match nothing else.

  • A spoken reply cannot exceed 8,000 characters.
  • Audio in a reprompt cannot run longer than 90 seconds in total.
  • A reprompt plays when the session stays open and the person says nothing that maps to an intent; they then have a few seconds to answer.
  • Certification checks policy, security, functionality and user experience.
  • Skills that allow purchases, and skills that handle protected health information, meet extra requirements.

Those limits come from Amazon’s request and response reference and its certification requirements. Designing the help, stop and fallback turns first, rather than last, is what keeps a skill from failing review on a missing intent.

Siri and App Intents on Apple devices

On Apple platforms, a product reaches Siri by exposing its actions and content through the App Intents framework. Apple describes app intents as the actions an app performs and app entities as its key data, which Siri, Spotlight, the Shortcuts app, widgets and the Action button can then use.

  • Name actions and content in the words people already use, such as track, song or podcast.
  • Expose the most popular actions first, and the contexts where they happen.
  • Offer content that is relevant to the person, such as recent items or favorites.
  • Refer to Siri by name, never with a pronoun.
  • Never impersonate Siri or use reserved phrases such as “Hey Siri” or “Call 911”.

The iPhone and iPad side of that build sits with our iOS app development team.

Google Assistant after Conversational Actions

Google ended Conversational Actions, its custom conversations for Assistant users, on June 13, 2023, according to its sunset overview. The same page says App Actions, smart home, Actions from web content and Media Actions were not affected.

For Android apps, Android Developers documents App Actions with built-in intents and shortcuts, supported on Android 5, API level 21, and higher. Google also notes that it decides when to surface an action, based on quality and relevance to the request. Our Android app development team builds that integration.

Does the end of Conversational Actions affect a new project?

It changes where the conversation lives. A brand that once planned a custom Assistant conversation now builds voice into its own Android app through App Actions, into an Alexa skill, into Siri through App Intents, or into its own app or phone line with its own speech stack.

Where a voice user interface can run, and the rules that apply
PlatformHow a product reaches itRules that shape the design
Amazon AlexaA custom skill built with the Alexa Skills KitInteraction model, required built-in intents, response limits, certification
Apple SiriApp Intents that expose an app’s actions and entities to Siri, Spotlight and ShortcutsSiri Human Interface Guidelines on responses, wording and reserved phrases
Google Assistant on AndroidApp Actions and the other paths Google lists as unaffected by the 2023 sunsetConversational Actions ended June 13, 2023
Your own app or phone lineSpeech recognition and synthesis inside the productYour own prompts, error paths and accessibility checks

Personas and sample dialogs before any build

Google recommends starting from clear requirements, a persona and sample dialogs. A persona is a design tool for writing consistent conversations. Sample dialogs give a quick sense of how the exchange sounds, without code or flow diagrams, and Google suggests role playing them with a second person, then listening to each system line in the text to speech voice it will use.

Google also advises designing for a screenless device first, because starting with a screen in mind makes it easy to lose the thread of the conversation.

  1. List the tasks people should finish by voice, in order of value.
  2. Write the persona in a few lines: tone, formality and what it never says.
  3. Draft sample dialogs for the happy path of each task.
  4. Add dialogs for corrections, silence, misses and system failures.
  5. Role play every dialog with two people, then play the system lines in the product’s voice.
  6. Turn the dialogs into intents, slots and prompts.
8,000: Characters per Alexa spoken reply. The ceiling on a skill's outputSpeech.
90 seconds: Audio allowed in a reprompt. Alexa's limit across reprompt audio files.
2: No Match attempts, then end. Google's guidance before closing a conversation.
June 13, 2023: Conversational Actions sunset. Google's end date for custom Assistant conversations.
Level A: WCAG 2.5.3, Label in Name. Visible labels match the words people speak.
Android 5: App Actions floor. API level 21 and higher, per Android Developers.

How is a voice UI tested with real speech?

It is tested out loud, in the places people will use it. Typed test phrases miss what speech brings: accents, background noise, false starts, fragments and people talking over the prompt.

  • Run sessions in a quiet room, a car and a kitchen.
  • Record which requests fell to the fallback intent, and add utterances for them.
  • Time each prompt read aloud at the product’s speaking rate.
  • Check every error path at least once per release.
  • Review the logs weekly in the first month after launch.

AI models that sit behind the voice layer are built by our AI agent development and AI integration services teams, and the wider choices about AI in a product are covered on our AI in software development page.

What voice interface work costs

The bands below run from the typical low to the published high in the US market. Design for a voice feature inside an existing app sits at the lower end; a full assistant that connects to calendars, CRMs and other systems sits higher.

Published US price bands (market ranges only)
WorkTypical low to published high
Voice interface design for an app, as UI/UX design work$8,000 to $40,000 per project
Voice and text assistant with CRM and calendar connections$30,000 to $80,000 per project
Custom voice or chat assistant build$30,000 to $150,000 per project
Horizontal range bars for voice interface design, a voice and text assistant with integrations, and a custom assistant build.Horizontal range bars for voice interface design, a voice and text assistant with integrations, and a custom assistant build.
Published US price bands (market ranges only) for voice interface design and voice assistant builds, typical low to published high.

Progression’s price for voice work depends on scope, is set in the written proposal and is subject to change. Model and voice usage, where an assistant uses them, is billed by volume and sits outside the build.

Pricing depends on scope and is confirmed in a written proposal. Published prices are subject to change.

How Progression designs and builds voice interfaces

Progression’s team takes voice user interface design from the first task list to a released product: sample dialogs, intents and slots, prompts, error paths, confirmations, screen handoff, accessibility checks, and the build on Alexa, Siri through App Intents, Android App Actions or the product’s own app and phone line.

What this page leaves out: the Progression Method, proprietary and reserved for clients. Reach out to learn more

Frequently asked questions

What does voice user interface design cover?
It covers what a voice product understands, says and does: the intents and their sample utterances, the slots for details, the prompts, the confirmations, the error paths and the handoff to a screen when one is present.
How is a voice user interface different from a graphical one?
It is heard rather than seen, so people cannot scan, scroll or look back. That makes prompt length, the number of choices read aloud and the recovery from misheard input the main design problems.
What is an intent in a voice app?
An intent is an action that fulfills a spoken request, such as booking an appointment. In Amazon’s terms it can carry slots, which hold details like a date or a city.
What are sample utterances?
They are the likely phrases people say for each intent, gathered as widely as possible so the system recognizes requests however they are worded.
How should a voice app handle a request it does not understand?
With a short reprompt first, then more help such as examples on the second attempt, and a graceful ending after two misses, which follows Google’s guidance on No Match errors.
What should happen when a person says nothing?
The app rephrases the question as if it was never heard and, if the person may be unsure, offers options. Google advises against saying “I didn’t hear you”.
When should a voice app ask for a yes or no confirmation?
Before actions that are hard to undo, such as a booking, a payment or deleting data, and when a misunderstanding would cost a lot. Most other details can be confirmed implicitly as the conversation moves on.
How long can an Alexa response be?
Up to 8,000 characters of speech, with reprompt audio limited to 90 seconds in total. Good prompts are far shorter than either limit.
Which intents must every Alexa skill handle?
AMAZON.StopIntent must be implemented, ending the session, and well built skills also handle AMAZON.HelpIntent, AMAZON.CancelIntent and AMAZON.FallbackIntent for requests that match nothing else.
Can Google Assistant still run custom conversations?
Custom Conversational Actions ended on June 13, 2023. Google says App Actions, smart home, Actions from web content and Media Actions were unaffected, so Android apps reach Assistant through App Actions.
How does an app work with Siri today?
Through the App Intents framework, which exposes the app’s actions and content to Siri, Spotlight, Shortcuts, widgets and the Action button.
What makes a good spoken prompt?
It is specific, short and complete without a screen. Apple’s examples include asking “Which soup?” rather than “Which one?” and leaving the app’s name out of replies.
How many options should a voice prompt read aloud?
Few enough to remember. When the full list is long, ask an open question that narrows it first, then offer a handful of choices.
Do voice interfaces help accessibility?
They help many people, provided they sit beside a non voice path and on-screen controls carry accessible names that contain their visible labels, as WCAG 2.5.3 asks.
Should a voice assistant have a persona?
It should have a defined one, written in a few lines, so tone and wording stay consistent across every prompt. Google treats the persona as a design tool for writing conversations.
Can one design serve Alexa, Siri and our own app?
Much of it can. Tasks, sample dialogs, prompts and error paths carry across, while each platform adds its own model, limits and review rules.
Which tests catch voice UI problems ahead of release?
Spoken sessions with real speakers in realistic places catch the most, along with a review of which requests fell through to fallback and a timing of every prompt at speaking pace.
Does a voice app need a screen design too?
Usually yes, because many devices now have displays. The spoken reply carries the essentials, and the screen adds detail without being required.
What does voice interface design cost?
Published US bands run $8,000 to $40,000 for voice interface design inside an app and $30,000 to $150,000 for a custom assistant build. Progression sets its price in the written proposal.
Is Alexa skill development a separate project from app work?
It can share the same tasks, dialogs and backend as the app, with the skill’s interaction model, intents and certification handled as their own stream of work.
Can Progression design and build our voice interface?
Yes. Progression’s team designs the conversation and builds it on Alexa, Siri, Android or the product’s own app and phone line, from New York for companies across the US, with on-site sessions on request.

Planning a voice feature?Tell us the task people should finish by voice and the devices they use. You get a free proposal with the intents and the first sample dialogs.

Get a free proposalTell us about a project you’d like help with.

Get a free marketing proposal

Tell us what you are trying to grow and we will come back with a plan, not a pitch deck. Same-day reply on weekdays.

Privacy Preferences
When you visit our website, it may store information through your browser from specific services, usually in form of cookies. Here you can change your privacy preferences. Please note that blocking some types of cookies may impact your experience on our website and the services we offer.
Contact Us