Designing a Voice: What Describe It and Muse Sounds Like That Means
Muse lets you customize its voice by describing it in words rather than picking from a preset list. That small design choice says more about Meta's agent strategy than it looks.
Buried in the feature list from Meta Connect 2026 is a detail easy to skip past: Muse lets users customize its voice "by describing what you want: faster, slower, or even a specific accent," according to Meta's Connect recap. No dropdown of five preset voices, no slider for pitch — you describe the voice you want in natural language, and Muse's real-time voice mode adapts to match.
That's a small UI decision with a large implication underneath it. It means the voice itself is generated, not selected — a continuous, describable space rather than a fixed menu — and it means Meta is treating voice personalization as important enough to build a natural-language interface for, rather than shipping a simple preset picker and calling it done.
Why voice, and why now
Muse's real-time voice mode is one of its most distinctive interaction patterns: users can carry on an extended spoken conversation with the agent while it continues completing tasks in the background, rather than the strict turn-taking of typed chat. That's a genuinely different mode of interacting with an AI agent than anything ChatGPT's Dots or xAI's Grok Bot have emphasized in their launches — both are described primarily as text/chat-driven, reachable through messaging surfaces like Slack or phone threads, rather than built around sustained live conversation as a first-class mode.
Meta's bet is plausible for a specific reason: an agent that's supposed to work continuously and check in with you when something changes benefits enormously from a communication channel that doesn't require you to stop and type. If Muse is meant to interrupt your day with "the flight price dropped, want me to rebook?" a five-second voice exchange while you're doing something else is a much lower-friction interaction than reading and typing a reply. Voice isn't decoration here — it's the delivery mechanism for the persistence and proactivity Meta is pitching as Muse's core differentiator, which we cover in What Is Muse?.
From voice to a face
Meta didn't stop at audio. Muse Realtime Avatar, developed by Meta Superintelligence Labs, is described as "state-of-the-art embodiment technology that turns Muse Realtime Voice into expressive, interactive avatars" — giving the voice a face, body, and set of expressions to go with it. Zuckerberg has reportedly given the default avatar a name, "Jolly," per TechCrunch's coverage of everything new coming to Muse. We look at what that avatar layer adds — and where it risks becoming decoration rather than function — in Muse Realtime Avatar: Giving a Voice Agent a Face.
A describable voice and an expressive avatar are both solving the same underlying problem: making a system capable of real autonomy feel legible and approachable enough that people will actually let it act on their behalf. Trust in an agent this capable has to be earned somewhere, and Meta is clearly betting some of that trust gets built through how the agent sounds and looks, not just what it does.
This is a pattern showing up across the entire agent launch cycle, not just at Meta. OpenAI gave Dots cartoonish "little floating" avatars that press coverage explicitly compared to Meta's approach — a comparison that shows up directly in reporting on both launches. Grok Bot, by contrast, has shipped with comparatively plain, enterprise-toned branding and no equivalent avatar push in the sources covering its launch, which tracks with its positioning toward business and developer workflows rather than a mass consumer audience. We take up that pattern directly in The Cartoon Problem, Again.
The unresolved question
What isn't yet clear from Meta's public materials is how far the "describe it" voice customization actually extends — whether it supports the full range of accents and speech patterns a user might request, how it handles ambiguous descriptions, and whether there are guardrails preventing the feature from being used to make Muse impersonate a specific real person's voice. None of that has been detailed publicly as of this launch, and it's the kind of feature where the interesting edge cases only surface once a large, adversarial user base starts actually pushing on it — which, given Muse's reported app-chart performance since launch, is already underway.
Why voice cloning risk is different for an agent than for a toy
It's worth being specific about why this matters more for Muse than it would for a novelty voice-changer app. A voice-customizable agent that also has real-world capabilities — the ability to send emails, make purchases through Link, and communicate with connected services on a user's behalf — means a voice-impersonation failure isn't just an awkward or unsettling audio clip. It's potentially the voice interface for something that can act with real consequences, which raises the stakes on getting the guardrails right well beyond what a standalone voice-generation tool would need to worry about. Meta hasn't published a policy specifically addressing voice-likeness requests, and that's a gap worth watching for, particularly given how central voice interaction is to Muse's whole design philosophy.
It's a useful contrast to how the rest of the industry is handling the same underlying capability. Voice cloning and description-based voice generation have been contentious features wherever they've shipped — largely because the same technology that lets you politely ask for "a calmer, slower voice" is technically adjacent to the technology that could be misused to approximate someone else's voice without consent. Meta shipping this feature inside an agent that already has demonstrated real-world purchasing power, rather than inside a standalone creative tool, raises the bar for how carefully the guardrails need to be built — and it's a detail that deserves more scrutiny from independent reporting than it's received in the launch-week coverage so far.
There's a milder, more everyday version of the same design question too: how much latitude "describe it" gives a user before the request stops being a reasonable stylistic preference and becomes something the system should decline. A request for "a warmer, more formal tone" is obviously fine. A request that starts inching toward mimicking a specific celebrity's cadence or a named individual's speech patterns sits in murkier territory, and it's the kind of boundary that tends to get defined by early users testing the edges of a feature rather than by a policy document published in advance. Meta hasn't described where that line sits for Muse, and it's a reasonable thing to watch for as more people actually use the customization feature at scale rather than in a controlled demo setting.
Part 5 of 30 in the Meta Muse series. Previously: Sentinel: The Agent That Approves Muse's Internet Access. Next: Checkout via Link: How Muse Actually Pays for Things. Browse more builder insights or explore AI skills at aiskill.market.