A new breed of voice AI startup is making waves, and Sesame AI is right in the middle of it. Co-founded by Oculus’ Brendan Iribe and Ubiquity6’s Ankit Kumar, the company is currently negotiating a funding round exceeding $200 million, with Sequoia Capital and Spark Capital reportedly leading the effort.
If closed, the round could push Sesame’s valuation north of $1 billion—anchoring the company as one of the most closely watched players in conversational AI.
What’s driving this surge of interest? Sesame’s answer isn’t more speed or more scale. It’s realism—an emotionally aware AI that doesn’t just sound smooth, but sounds alive.
Flawed by Design: A Voice That Stumbles, Pauses, and Feels Real
Sesame’s Conversational Speech Model (CSM) sits at the core of its product. Powering digital assistants named Maya and Miles, the model embraces imperfections like hesitations, stutters, tonal shifts, and inconsistent pacing. This isn’t a bug; it’s intentional. Users described their experience with the voice assistant as “eerily human-sounding” and even “uncomfortable.”
The assistant doesn’t just mimic tone. It interprets emotional signals in the user’s voice—shifting to a slower, more soothing tone when stress is detected, or becoming playful during creative interactions. The assistant can role-play, adjust to character prompts, and shift demeanor depending on context. It’s designed to react in real-time to the shape and rhythm of a conversation, not simply the words spoken.
As explained in Sesame’s official research publication, “Crossing the Uncanny Valley of Voice”, the model dynamically shifts its delivery based on contextual signals. This allows the AI to respond in ways that feel more emotionally authentic rather than mechanically pre-scripted.
Open-Source Model, Hardware Plans, and a Hugging Face Demo
Sesame has released its CSM-1B model on GitHub under the permissive Apache 2.0 license, opening the door for developers to build on it with minimal restrictions. The 1B parameter base model can also be tested directly via a hosted demo on Hugging Face.
The architecture relies on Residual Vector Quantization (RVQ), a technique that compresses audio inputs into efficient token sequences. CSM processes these alongside textual data, making it capable of responding with contextually aware, emotionally tuned speech.
While it currently avoids dependency on large pretrained language models, Sesame has outlined plans to integrate such systems and expand language support to over 20 languages in future iterations.
In parallel, Sesame is developing lightweight AR eyewear designed for everyday use. Unlike visually immersive headsets, the product is focused on audio and offers all-day interaction with its AI assistant.
Anjney Midha, general partner at Andreessen Horowitz, one of Sesame’s earliest investors, wrote in a blog post this February: “Sesame is built around the simple, but non-obvious, idea that the answer isn’t in the screens of AR glasses — it’s in the audio. To date, the emotional flatness of AI audio has been exhausting and unnatural. But if you remove the visual display from AR glasses and instead focus on an amazing audio-first AI system, you can create a computing experience that feels seamless and intuitive.”
Funding Momentum and Strategic Backers
Sesame’s funding round isn’t just attracting Sequoia, Spark and Andreessen Horowitz. It also counts Matrix Partners among its backers. The company’s leadership combines Iribe’s experience in hardware platforms like Oculus with Kumar’s background in spatial computing and Discord’s community architecture—giving it technical depth and real-world product intuition.
The pitch to investors is clear: build the operating system for voice-first computing. Rather than challenge OpenAI and Google on speed or scale, Sesame is leaning into expressivity, nuance, and persistent presence. It’s less Alexa, more ambient companion.
Industry Context: Expressive Voice AI Heats Up
Sesame is not operating in a vacuum. Big tech is converging fast on expressive voice. OpenAI’s Advanced Voice Mode, rolled out to the web in late March, introduced better turn-taking and latency reductions.
It avoids interrupting users during pauses and has begun tweaking personality traits to create a more interactive experience. That feature remains gated behind premium tiers, though OpenAI expanded access in February 2025 to free users with limitations.
Google’s Chirp 3 model, integrated into Vertex AI, offers Instant Custom Voice tools and expressive tone controls across 31 languages. It emphasizes personal branding, call center support, and localization—approaches that contrast with Sesame’s focus on emotional authenticity. Chirp 3 also highlights ethical challenges, particularly around voice cloning and data consent, which could surface for Sesame as well.
Microsoft’s Copilot assistant, which now features freely available voice interaction, rounds out a fast-evolving competitive landscape. Meanwhile, other AI projects—like “Unhinged” Grok mode from Elon Musk’s xAI—are exploring expressive speech in more extreme directions.
Emotional Intelligence, Risk, and Real-World Friction
As the technology improves, so do concerns around deception and misuse. Sesame’s assistant doesn’t impersonate real people, but its realism blurs lines in human-machine interaction.
This realism also poses design and performance trade-offs. Running emotionally responsive models in real time, especially on wearable devices, comes with high compute costs. Processing natural dialogue on-device requires power-efficient chips and low-latency architecture—areas that Sesame has yet to detail publicly.
The company’s emphasis on realism could put strain on battery life or thermal limits in hardware form factors like glasses.
Despite those hurdles, the interest around Sesame is growing. Between open-source releases, ambitious hardware integration, and a valuation reportedly crossing the billion-dollar mark, the startup is staking a claim not just on how AI sounds—but on how it feels to talk to one.
