Should your AI sound human — or like you?What to know before designing a voice agent.A few months ago, I received a voicemail that made me pause. It was warm, clear, and casually confident: “Hey there! Just checking in to see how you’re doing and if you need anything.” I thought it was a friend. It wasn’t. It was a voice AI.​For a split second, I didn’t know how to feel. That uncanny chill — it wasn’t fear. It was disorientation. Like meeting someone you almost remember, but not quite. Like hearing your own laugh from a stranger’s mouth.​This wasn’t just the uncanny valley. This was the uncanny voice.We Know What the Uncanny Valley Looks Like. But What Does It Sound Like?​The uncanny valley, classically, refers to robots or animations that look almost human — but not quite.Source Uncanny Valley by Anika H. ‘26Their near-humanness makes them eerie. Voice, it turns out, has its own version of this. Not just in how words are pronounced, but how emotions are expressed. Too perfect, and it feels fake. Too flat, and it feels dumb. Somewhere in between is where we get that gut-level reaction: this is not okay.​As voice AI evolves, more companies are confronting this invisible threshold. Do you want your voice agent to sound real? How real? Real like a receptionist? Real like you?​Let’s back up.What Voice Agents Are Actually Doing​Voice AI isn’t just about speech synthesis. It’s an identity layer. Every time a customer hears that voice, they’re forming an impression — of your product, your professionalism, your trustworthiness. So, what persona do you give that voice?​a16z AI Voice Agents: 2025 UpdateHere are the common approaches:The Owner CloneA voice that mimics the founder or owner. It feels personal, familiar, and trustworthy — especially for small businesses where the owner’s voice is the brand. Think boutique fitness studios, clinics, or neighborhood cafes. It works best when the owner is already the face (and voice) of the brand.Such as: Be My Eyes, a visual assistance app for the blind, now integrates voice AI where the founder’s tone and phrasing were studied to train the assistant — maintaining trust and familiarity for users used to hearing from the original human volunteers.The Brand Proxy: A crafted voice that sounds like the brand. Think luxury? It’s calm and elegant. Think fast food? It’s upbeat and fun. This voice is scripted, styled, and tuned to the brand’s identity — it doesn’t matter who is speaking, only how.Such as: Taco Bell uses an upbeat, slang-savvy voice in its AI drive-thru systems to keep the tone light and playful — matching the youthful, quirky spirit of the brand.The Clear Robot: It owns its AIness. “Hi, this is an automated assistant.” No pretense, no confusion. It’s designed to be efficient, transparent, and unobtrusive. It doesn’t try to be human — it tries to be useful.Such as: IBM Watsonx Orchestrate delivers clear, robotic voice agents for customer service, especially in complex B2B environments where trust is built through clarity, not charm.Each has tradeoffs. But the big question is: Should you be upfront that it’s AI, or try to pass for human?What the Research Actually SaysMultiple studies confirm what our instincts already hint at: people trust AI voices more when they know they’re artificial, but they prefer ones that sound natural, warm, and expressive.A study by researchers at the University of Tokyo explored public attitudes toward AI ethics and found that transparency significantly shapes user trust and perception of fairness — especially when users understand they’re interacting with an AI system. When AI identity is disclosed, people tend to recalibrate their expectations and evaluate the system more thoughtfully.At the same time, researchers found that users responded more positively to voice assistants that exhibited human-like social cues — such as prosody, conversational timing, and vocal warmth. These cues enhanced the perceived social presence of the assistant, leading to increased user satisfaction and trust.This study highlights the importance of incorporating human-like social cues in voice assistants to improve user experience and satisfaction.So we’re walking a tightrope:Too robotic = users disengage.Too human-like without disclosure = users feel tricked or uncomfortable.The sweet spot?✨ Sound natural and expressive — without pretending to be human.That’s the balance top players are trying to strike.OpenAI’s Voice Engine and ElevenLabs’ cloning tech can now replicate a human voice from just 15 seconds of audio. That’s not future-speak — that’s happening now (OpenAI example, ElevenLabs). In internal tests, the clones are indistinguishable from the real speaker in blind trials.https://medium.com/media/0d51b4761e7da902effaacde8b6400b6/hrefThis unlocks some powerful use cases — but it also raises ethical red flags. When you can recreate someone’s voice this accurately, restraint isn’t just recommended — it’s required.What Happens When It Sounds Too Real?​Imagine your therapist’s voice calls you. It remembers your last session. It uses your name. It checks in with warmth and pauses just long enough to feel like it’s truly listening. But it’s not your therapist. It’s an AI — a voice model trained to sound just like them.Startups like Sesame AI are closing the gap between machine and human with startling precision. Their voices carry presence — not just sound, but the illusion of being there with you. And presence, when faked, becomes performance.You didn’t consent to that. And even if you did, it still feels off.Because the more human a voice becomes, the more responsibility it carries. We don’t want a tool pretending to care. We want it to know it’s a tool. Or at the very least — we want to know.https://medium.com/media/2241a83b7d28f8d6e183dfeaa7960b4b/hrefSo What Should We Do?​Transparency wins trust.Always let people know they’re talking to AI.Example: Replika introduces itself as an AI companion and uses visual cues and tone to clearly differentiate itself from a human. Users build trust because they know what they’re interacting with.Familiarity builds connection.Voices that match the tone, rhythm, and energy of your brand feel more coherent.Example: Duolingo’s voice assistant uses playful, expressive voices that mirror its brand personality — casual, quirky, and encouraging. The voice reinforces the entire learning experience.Slight imperfections help.Just like design embraces whitespace, voice AI should embrace subtle pauses, slight hesitations, and emotional range.Example: Sesame AI intentionally incorporates “voice presence” — pacing, prosody, and even momentary silence — to create interactions that feel authentic without being deceptive.And most importantly: voice should feel like a choice.Let people opt into voice.Example: Google Assistant is available across platforms but always gives users a fallback — tap, type, or ignore. No pressure. That autonomy builds loyalty.And perhaps most importantly: voice should feel like a choice.​If your customer would rather text, let them. If they like voice, make it a delightful one.​Where This Is GoingIn a few years, most small businesses will have voice agents by default. It’ll be weird if they don’t. The winners won’t be the ones who sound the most human. They’ll be the ones who sound the most intentional.Not perfect. Not indistinguishable from people. Just clear about who they are, consistent with the brand they represent, and conscious of the emotional space they’re entering every time they speak.Because the future of voice isn’t about passing the Turing Test. It’s about earning trust — one word at a time.References A Mixed-Methods Approach to Understanding User Trust after Voice Assistant FailuresTrust in Virtual Agents: Exploring the Role of Stylization and VoiceShould your AI sound human — or like you? was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.