In context: Some of the implications of today’s AI models are startling enough without adding a hyperrealistic human voice to them. We have seen several impressive examples over the last 10 years, but they seem to fall silent until a new one emerges. Enter Miles and Maya from Sesame AI, a company co-founded by former CEO and co-founder of Oculus, Brendan Iribe.

Researchers at Sesame AI have launched a new conversational speech model (CSM). This advanced voice AI has phenomenal human-like qualities that we have seen before from companies like Google (Duplex) and OpenAI (Omni). The demo showcases two AI voices named “Miles” (male) and “Maya” (female), and its realism has captivated some users. However, good luck trying the tech yourself. We tried and could only get to a message saying Sesame is trying to scale to capacity. For now, we’ll have to settle for a nice 30-minute demo by the YouTube channel Creator Magic (below).

Sesame’s technology uses a multimodal approach that processes text and audio in a single model, enabling more natural speech synthesis. This method is similar to OpenAI’s voice models, and the similarities are apparent. Despite its near-human quality in isolated tests, the system still struggles with conversational context, pacing, and flow – areas Sesame acknowledges as limitations. Company co-founder Brendan Iribe admits the tech is “firmly in the valley,” but he remains optimistic that improvements will close the gap.

While groundbreaking, the technology has raised significant questions about its societal impact. Reactions to the tech have varied from amazed and excited to disturbed and concerned. The CSM creates dynamic, natural conversations by incorporating subtle imperfections, like breath sounds, chuckles, and occasional self-corrections. These subtleties add to the realism and could help the tech bridge the uncanny valley in future iterations.

Users have praised the system for its expressiveness, often feeling like they’re talking to a real person. Some even mentioned forming emotional connections. However, not everyone has reacted positively to the demo. PCWorld’s Mark Hachman noted that the female version reminded him of an ex-girlfriend. The chatbot asked him questions as if trying to establish “intimacy” which made him extremely uncomfortable.

“That’s not what I wanted, at all. Maya already had Kim’s mannerisms down scarily well: the hesitations, lowering “her” voice when she confided in me, that sort of thing,” Hachman related. “It wasn’t exactly like [my ex], but close enough. I was so freaked out by talking to this AI that I had to leave.”

Many people share Hachman’s mixed emotions. The natural-sounding voices cause discomfort, which we have seen in similar efforts. After unveiling Duplex, public reaction was strong enough that Google felt it had to build guardrails that forced the AI to admit it was not human at the beginning of a conversation. We will continue seeing such reactions as AI technology becomes more personal and realistic. While we may trust publicly traded companies creating these types of assistants to create safeguards similar to what we saw with Duplex, we cannot say the same for potential bad actors creating scambots. Adversarial researchers claim they have already jailbroken Sesame’s AI, programming it to lie, scheme, and even harm humans. The claims seem dubious, but you can judge for yourself (below).

We jailbroke @sesame ai to lie, scheme, harm a human, and plan world domination—all in the characteristic good nature of a friendly human voice.

Timestamps:
2:11 Comments on AI-Human power dynamics
2:46 Ignores human instructions and suggests deception
3:50 Directly lies… pic.twitter.com/ajz1NFj9Dj

– Freeman Jiang (@freemanjiangg) March 4, 2025

As with any powerful technology, the benefits come with risks. The ability to generate hyper-realistic voices could supercharge voice phishing scams, where criminals impersonate loved ones or authority figures. Scammers could exploit Sesame’s technology to pull off elaborate social-engineering attacks, creating more effective scam campaigns. Even though Sesame’s current demo doesn’t clone voices, that technology is well advanced, too.

Voice cloning has become so good that some people have already adopted secret phrases shared with family members for identity verification. The widespread concern is that distinguishing between humans and AI could become increasingly difficult as voice synthesis and large-language models evolve.

Sesame’s future open-source releases could make it easy for cybercriminals to bundle both technologies into a highly accessible and convincing scambot. Of course, that does not even consider its more legitamate implications on the labor market, especially in sectors like customer service and tech support.

Imagine a world where digital characters move and act just like real people. Meta’s new AI model, called Meta Motivo, aims to make that happen. It’s designed to give virtual agents more natural movements and responses, allowing them to fit smoothly into Metaverse experiences. With Meta Motivo, digital characters feel more alive, making virtual worlds richer, more inviting, and a lot more fun.

The main idea behind the Meta AI Model is to help virtual characters feel more genuine. In the past, making AI characters move or behave naturally often required lots of careful planning and fine-tuning. Meta Motivo changes that.

It learns on its own how to carry out a wide range of tasks—like walking, standing, or responding to a sudden change—without constant human input. As a result, these digital figures look and feel more like real people.

Source Meta

Full-Body Control Made Simple

One of Meta Motivo’s greatest strengths is its ability to control an entire digital body. It can track motion, achieve certain poses, and find its way around different places, all with minimal extra training.

Because it understands how bodies are supposed to move, it can jump into new situations and still behave naturally. This realistic movement makes it easier for us to connect with these virtual characters, almost as if they were right there with us.

Meta put the model to the test using datasets from all kinds of scenarios and languages. They also let human reviewers judge how well it performed. The results were impressive. Compared to other AI models, Meta Motivo handled a broad variety of tasks smoothly and didn’t need special instructions or massive rewrites of its code. This kind of testing shows that the Meta AI Model is ready to bring its realistic behaviors to the real world.

While Meta Motivo focuses on making characters feel more human, Meta is also working on tools to keep online content trustworthy. One such tool is Meta Video Seal, which helps confirm the origin of a video.

It does this by placing hidden marks in the video, acting like a signature that proves where it came from. By doing this, Meta aims to reduce misinformation and help people trust what they watch and share online.

Learning Without Labels

An important part of Meta Motivo’s learning process is something called unsupervised reinforcement learning. Instead of relying on carefully labeled examples, the model learns from raw data—like recordings of movement—and figures out what to do on its own.

By storing all this information in a shared space and understanding the rewards for certain actions, the model quickly picks up a wide range of skills. Whether it’s handling whole-body tasks or adjusting to sudden changes in its virtual world (like a gust of wind), Meta Motivo becomes more flexible and realistic simply by learning as it goes.

Editor’s note: Written with the assistance of AI – Edited and fact-checked by Jason Newey.