We have Her. Why aren't we using Her?

It’s already been two years since Sam Altman first tweeted “her” and kicked off the race toward what has become the “holy grail” of voice AI.

Now, in 2026, you can’t throw a rock in the Valley without hitting five startups claiming they’re building Her. The craziest thing is how good the tech has gotten - in a way, we’re already there.

Gemini Live, GPT Realtime 2, and Sesame are just a few of the agents launched this year with the ability to express human-like prosody and emotion, with minimal latency, multi-modality, and tool calling. While the rumored GPT-BiDi-1/Live, and Thinking Machines’ upcoming full-duplex voice model are just around the corner.

For just $20/month (less than what I spend on DoorDash in a day) you get a voice AI you can confide in, flirt with, and even use to “manage your calendar”.

And yet…how many people do you know who actually use voice AI every day?

Sure I might use it if I’m driving and need to answer a question hands-free. But these occasions are few and far between, and I don’t find myself using it any other time.

How is it possible that we’ve had so many dramatic improvements to voice AI in recent years, yet consumer behavior has still barely shifted?

We have Her. Why aren’t we using Her?

1Consumers Just Need Time to Catch Up

We’ll start with the most optimistic case: the tech is here, but consumer behavior just takes time to adjust.

A classic example is the iPhone, which first launched in 2007 but only sold about 1.4 million units that year. It wouldn’t be until the end of the second year that Apple finally cleared its original 10 million-unit goal. Really good voice AI is only just starting to come out this year, so we may still be in that initial lull.

If we’re looking at a Her-like experience more broadly however, there may already be signs of people “living in the future.” AI companion apps like Character AI attracted millions of users wanting to chat with their favorite anime characters and influencers - with many of them even developing deep romantic attachments. It’s even gotten to the point where Open AI had to roll back part of their update to GPT-4o last year when users loudly decried the loss of their favorite companions.

While this behavior points in the direction of Her, it’s also quite damning in that voice was not essential to making these products sticky. In fact, the roll out of voice features in these companion apps came relatively uncelebrated, and most AI companion power users still prefer text to this day.

It may be that we need to see an AI companion app go voice-native - enter Sesame.

Founded by the former head of Oculus, Sesame made headlines last year for showing off the most natural-sounding voice agent ever (to this day I still don’t think they’ve been beat). They accomplished this using today’s technology - no fancy new models, just careful, intentional efforts to improve prosody, speech fillers, and gathering high quality recording data. The result is something that literally feels like you’re talking to Her (or Him - they had a male persona too).

This past month saw the release of their iOS and Android apps, which come with additional features like tool calling, deep research, and more personas to talk to. For my money it’s the best consumer voice experience out there bar none - and it confuses me to no end why this isn’t all over the headlines.

Even as I say all of this however, I still found myself dropping off after just three days of use.

As impressive as the tech is, it only took a few calls before the conversation patterns started feeling repetitive. The speech fillers became predictable, the affectations got annoying, and the illusion fell apart - at which point it was just a strictly less efficient Chat-GPT.

I struggled to find situations where I actually wanted to use it.

On the one hand, maybe I’m simply not the target user and it’s only a matter of time before a Sesame style voice app achieves hockey stick growth with a consumer audience.

On the other hand, maybe there’s still something fundamentally missing.

2We Don’t Actually Have Her Yet

The second perspective is more skeptical: not only do we not have Her, but depending on who you ask, we might not even be close.

Samantha, the AI in the movie, is not just a pretty voice with good prosody. She also knows when to speak and when to stay quiet. She adapts to subtle emotional shifts. She understands what Theodore is physically doing. She can read his writing, his emails, and listen to his environment.

Compared to where we’re at with current voice AI, there’s still 3 crucial pillars of research to solve:

1. Full-duplex conversation

Today’s voice AI follows something called discrete turn-taking:

You speak. The system detects that you stopped speaking. It thinks. Then it replies. The entire conversation happens within one channel (aka half-duplex).

Real human-to-human conversations however, have two channels of communication (aka full-duplex) - people might be speaking over one another, they might be back-channeling (“mhm”, “right”, “I see”), or even interrupting one another mid-dialogue. It’s the difference between a walkie-talkie and a phone call.

2024 saw a huge breakthrough with Kyutai’s Moshi which delivered the first ever real-time full-duplex speech model (and open sourced it no less!). Respect. Our team did an in-depth breakdown of the paper in our previous blogpost.

Nowadays, every major frontier lab in the world is pushing for full-duplex voice. In May we saw Thinking Machines announce their interaction models (which they claim can continuously take in audio, video and text while also actively thinking, responding, and acting in real-time), Open AI’s new GPT Live-1 model shows similar capabilities, as with Gemini Live, Nova Sonic, and many others. 2026-2027 may be the year of duplex voice AI.

Despite these exciting advancements however, it’s worth calling out that full-duplex isn’t a silver bullet. It currently comes with trade-offs in intelligence, tool-calling accuracy, and increased inference cost. This means for most consumer products it may not be worth the cost. We’ll have to see if any exciting new breakthroughs later this year will challenge this notion!

2. Understanding emotion

Most voice API services like Elevenlabs already feature “emotion tags” that allow you to generate TTS with inflections that make it sound happy, sad, excited, frustrated, etc. But sounding emotional isn’t the same as understanding emotion.

For years, speech emotion recognition looked much like image classification: a model would assign one of a handful of labels - happy, sad, angry, fearful, neutral. While useful, this is a massive oversimplification of how humans actually experience emotion.

More recently, rather than attributing a single, limited label (ie. “was this speaker angry?”), researchers are developing speech models that can learn latent representations of emotion directly from audio (emotion2vec is a good example of this from 2024). Latent representations allow us to encode the speaker’s emotional state directly. This means we can map all kinds of conversational subtleties - be it a person’s confidence, hesitation, sarcasm, speaking style, or countless other signals that don’t fit neatly into English words.

This raises another interesting question: how does one get those rich emotional representations into the language model? One popular approach, used by systems like Ultravox, is to feed audio through a pre-trained speech encoder (such as Whisper), project those embeddings into the language model’s embedding space with a small adapter network, and let the LLM reason over them alongside text. In principle, these embeddings should preserve information about prosody, tone, hesitation, and other acoustic cues that transcripts alone discard. In practice, however, it’s still an open question as to how much of this emotional nuance survives the projection and whether today’s language models are actually capable of exploiting it.

As voice models improve, we’re likely to see more and more of them adopt this technique (or similar ones), creating a much more Her-like ability to empathize.

3. World context

The previous two pillars we covered focus on improving speech, but to truly achieve Her we need an AI that can also understand the context of the “world” the conversation is happening in. After all, in the movie, Samantha is more than just a voice, she’s technically the entire operating system. She’s able to see through Theodore’s cameras, read through his emails, and even join him in games.

The answer for this may be found in world models - which I recognize is a very loaded, fuzzy term, but bear with me. World models learn the structure of an environment and can predict how the world changes over time.

This idea has become increasingly influential in AI research. Yann LeCun has argued that LLMs aren’t enough, true AI requires learning predictive models of the world, leading to approaches like Joint Embedding Predictive Architectures (JEPA), which learn abstract representations of environments instead of simply predicting pixels or tokens.

Games are constrained virtual worlds. Unlike the real world, in a game we can fully expose its state. This could allow the model to know where you are, what you’re carrying, what quest you’re on, what actions are possible - basically everything. Instead of inferring reality through sparse inputs, we can feed the model as much data as it needs. This makes games an unusually clean training ground for world-aware AI.

We’re not the only ones who believe in games. General Intuition recently released MIRA, a world model that simulates an entire multiplayer game of Rocket League. There is no hand-written physics engine or game logic - the model learns the dynamics of the world directly from thousands of hours of play data. It’s by far the most compelling gameplay we’ve seen from a world model to date.

Now imagine combining those two systems.

• A voice model that gives the AI the ability to communicate naturally (duplex and emotion).
• A world model that gives it the ability to understand what’s actually happening.

Instead of a chatbot, we’d have a kind of “co-presence”. And it may be the case that this is the level we need to achieve before we can truly say we have Her.

3Voice Calling isn’t actually the Right Interface

The third and final case is the most bearish - what if a Her-like “voice call” is fundamentally the wrong way to think about the consumer experience?

Take VR for example - it looks great in a movie, it sounds great in science fiction, and it feels amazing when you demo it. “Surely that’s the future,” I thought.

~$200 billion dollars and a decade later however, it turns out that the activation energy required for VR makes it too cumbersome to use. You have to put on a headset. You lose access to your phone. You can’t multitask. You need space. The device has to be charged. Your hair gets messed up. Sometimes you get motion sickness. It’s just easier to use a laptop for most things.

Her-like voice AI may have a similar problem - it looks great in a movie, but in real life it just makes things actively worse.

You cannot comfortably talk to your AI at your office. You cannot ask it sensitive questions while riding on the subway. You cannot even skim a long voice message and decide whether to commit more time to it. As a future standard for human-computer interaction, voice is just too linear compared to a screen.

It’s worth noting that Her wasn’t trying to lay out some grand vision for the future of voice AI. It was a commentary on the present: people beginning to “fall in love” with their phones through para-social relationships and social media (already true back in 2013, even more true now in 2026). Samantha was never meant to be a serious product design proposal.

So what does that leave for voice AI? If the “phone call with an AI” isn’t necessarily the endgame, where can voice actually fit?

We may already be seeing compelling consumer(ish) products emerge:

• Wispr Flow treats voice as a faster, more natural keyboard.
• Granola acts as a passive listener, taking meeting notes in the background.

These classes of products are a far cry from the vision presented in Her. Their success comes not from replacing human conversation, but from removing friction in situations where typing is the bottleneck. So where else are we feeling this kind of friction? Where else do we need to be hands-free?

Earlier I gave the example of speaking to my voice agent during a commute. The other obvious example that comes to mind is gaming. Discord became one of the most successful communication platforms in the world by embracing voice-first interaction, and interestingly, has struggled to expand far beyond the core gaming audience. It turns out people don’t necessarily want to talk all the time - they want to talk when talking is the most convenient option.

In fact, since texting became ubiquitous, I avoid phone calls whenever I can. Phone calls are, in many ways, a poor substitute for being together in person. Back when I was working in VR, we spent billions of dollars to create VR telepresence, but the tech is still a far cry from actually feeling the presence of another person in the room.

That makes me wonder whether we’ve been framing consumer voice AI incorrectly from the start. Instead of being the OS, voice may be at its best in the moments where screens and keyboards aren’t convenient: driving, walking, cooking, exercising, gaming. It’s probably not the universal interface that touchscreens became - but it may end up being the best interface whenever your hands are busy.

Conclusion

When I started writing this post, I thought I was trying to answer a simple question:

If voice AI has gotten so good, why isn’t everyone using it?

By the end, I don’t think there’s a single answer. Consumer adoption may simply take time. The technology itself still has quite a few gaps to cover. Or perhaps all of these companies have been chasing the wrong interface entirely. As cliche as it is, my guess is that all three cases are true to some extent.

I referenced the delayed success of the iPhone earlier in this post, and I’m reminded of it now. When it comes to the reason why smartphones eventually took over, people often point to the touchscreen, the industrial design, or the falling price point. But arguably the biggest inflection point was the invention of the App Store. Suddenly the iPhone stopped being just an impressive piece of hardware and became a platform that let people accomplish things they couldn’t before.

I keep coming back to the same question we heard over and over on our own company’s journey building voice AI characters:

“What does it do?”

That question is still largely unanswered.

It probably won’t be shaving 50 milliseconds off latency or increasing expressiveness by 10% that will bring about the future of consumer voice AI. Rather, it’ll be because people find out they can do something that’s only possible because of voice.

Maybe that’s a companion that inhabits your favorite game with you. Maybe it’s an assistant that follows you through your day while your hands are pre-occupied. Or maybe it’s something none of us have imagined yet.

Whatever it is, I suspect the killer feature won’t be Her. It’ll be what talking to a Her enables.

That’s the problem we’re interested in at Frisson Labs. We’re building the voice-world model that can answer the question of “what does it do”, and we think games are one of the first places where voice becomes the natural interface instead of a gimmick.

If this article resonated with you - or you think I’m completely off my knocker - I’d love to hear from you. Feel free to reach out.