AI Voice Choice: Give Users Control and Continuity

Imagine using an AI assistant to review a long report. The voice is clear, but its pace feels tiring. You want a calmer delivery without abandoning the task, losing your place or explaining your preferences again. That small interaction raises a larger design question: how much control should people have over the voice accompanying their work?

Voice personalisation can be treated as an occasional cosmetic choice. A more useful approach considers intelligibility, comfort, accessibility and continuity together. The person listening should be able to preview an available voice, change it and understand what the application will remember.

This article proposes continuity of voice preference as a design principle: preserve the user’s chosen characteristics where possible, while being transparent when the available voices change. It is a product recommendation, not a claim that an exact synthetic voice can remain available forever.

What voice research can tell us

A study by Kylie Goodman and Christopher Mayhorn, published in Applied Ergonomics in January 2023, examined perceived voice gender and pitch during a medication-information task. Voices perceived as female received higher trust ratings, while pitch did not significantly affect trust. The authors also found that trust and reliance primarily varied between participants; the manipulated vocal characteristics explained only a small share of within-person variation.

That specific experiment does not establish the best voice for every assistant, task or listener. Nor does it demonstrate that offering voice choice produces a lasting wellbeing benefit. It does illustrate why designers should examine the listener’s experience rather than assume that one vocal characteristic determines trust.

Preferences should therefore be gathered directly. A person’s preferred voice is useful information about their experience, without being a basis for inferring their personality, competence or relationship needs.

Offer a preview that supports a real comparison

Let users hear a short, representative passage before choosing. Use the same passage at a comparable volume across options so that a livelier script or louder recording does not dominate the comparison.

Describe characteristics that help people choose: language, pronunciation, pace, expressiveness and vocal style. Masculine-sounding, feminine-sounding and more gender-ambiguous descriptions may be useful where supported, but none captures everything a listener hears. Previews let people decide for themselves.

Avoid assigning capabilities to a voice label. A reassuring delivery does not demonstrate better reasoning, and an authoritative sound does not establish subject expertise. Test the quality of the answer separately from the presentation.

Make changing the voice part of ordinary use

A preference may change with the task or setting. Someone might want a measured delivery for reviewing instructions, a more expressive style for a story and text alone while travelling. The design should make those changes straightforward.

Where the technology permits, switching voices during a session should preserve the working conversation and the user’s place. If switching interrupts playback or requires a restart, explain that limitation before the change. Avoid forcing people to discover the consequences through trial and error.

Make the selected option visible and allow it to be changed again. A person who experiments with another voice should have an understandable route back to the previous available choice.

Remember the preference at the scope the user chooses

A saved voice setting can apply to the application, a particular project or a device. These are different choices. Explain the scope instead of leaving people to infer whether a selection will follow them into another workspace.

For example, a reader may want one voice for a fiction project while keeping a different default for work. Someone sharing a device may prefer a temporary setting. Offer only the scopes the application can actually support, and make the current one clear.

This sits alongside the broader personalisation questions in Your Personal AI Needs a Constitution: which preferences guide the assistant, which boundaries remain important and what the user can revise.

Preserve the preference when a provider changes

An application may depend on voices supplied by another service. Product design should account for the possibility that a voice becomes unavailable or changes. A promise of permanent access to one exact voice would need a basis the application can honour.

Consider a fictional learning application whose user has selected a calm, masculine-sounding voice with clear pronunciation. When that specific voice becomes unavailable, the application retains the general preference, explains the change and offers previews of available alternatives. It asks the user to choose rather than silently declaring that a replacement is equivalent.

The listener may find that none is suitable. Provide an available alternative such as text while making the limitation clear. Continuity is stronger when people understand the change and retain control over their response to it.

Keep basic audio controls easy to reach

Voice selection should accompany accessible controls for starting, pausing, stopping and adjusting playback. Consider a text alternative and usable controls for people navigating with a keyboard or assistive technology. These are design questions to test with the intended users.

W3C’s explanation of WCAG 2.2 Audio Control addresses webpage audio that starts automatically and lasts more than three seconds: users need a way to pause or stop it, or control its volume independently of system volume. The guidance explains that competing sound can interfere with screen-reader use.

That criterion concerns a particular autoplay situation; it should not be presented as a complete specification for conversational AI. The broader recommendation here is to make audio controllable and avoid making sound the only usable route through the task.

Test whether the choices improve the experience

Before treating voice personalisation as complete, observe people using it during realistic tasks. A useful review asks:

• Can people find and compare the available voices without losing their current work?
• Do the descriptions match what listeners actually hear?
• Can they switch, pause or return to text when their situation changes?
• Do they understand where their preference is saved?
• If a voice changes, can they recognise what happened and choose an acceptable alternative?

Collect feedback on intelligibility, comfort and task completion separately. A voice may be enjoyable while making a long passage harder to follow. Another may be easy to understand but unsuitable for the listener’s preferred use. Testing should reveal those differences rather than reduce the experience to one satisfaction score.

Introduce changes in a bounded prototype and state which features are proposals, which are implemented and which depend on a provider. Do not advertise persistent preferences or seamless switching until they work under the conditions being promised.

A thoughtful voice experience gives people an understandable choice, a way to change it and an honest explanation when continuity has limits. Those are concrete design commitments that can be checked.

Sources

• Goodman, K. L. and Mayhorn, C. B. (2023). It’s not what you say but how you say it: Examining the influence of perceived voice assistant gender and pitch on trust and reliance. Applied Ergonomics, 106, 103864. Published online 22 August 2022.
• W3C Web Accessibility Initiative. Understanding Success Criterion 1.4.2: Audio Control, explanatory guidance for WCAG 2.2, updated 16 September 2025.