Inworld, the Mountain View company building AI characters for games and interactive experiences, has acquired Ultravox, the Seattle voice AI startup that launched in 2022 as Fixie. The deal, announced October 2, was reported by GeekWire. Financial terms were not disclosed.
Ultravox was founded by Google and Apple veterans — Hessam Bagherinezhad, Matt Welsh, Justin Uberti and Zach Koch — to build a speech model for real-time, natural-sounding AI voice agents. The company raised a $17 million seed round in March 2023 led by Redpoint Ventures, which PitchBook valued at $51.46 million at the time, with participation from Madrona Venture Group, Zetta Venture Partners, SignalFire, Bloomberg Beta and Kearny Jackson, plus current and former executives at Google, Amazon and Apple.
Why Inworld wants it
Inworld’s business is AI non-player characters: game characters you can actually talk to. The company has raised $119.8 million since launching in 2021 and is valued at $515 million, according to PitchBook. Its characters already talk — what they don’t do well is the thing humans find effortless and machines find brutally hard: managing the flow of a spoken conversation.
That’s Ultravox’s specialty. The startup’s technology handles turn-taking and interruptions in voice conversations with AI — knowing when the human has finished speaking, when they’re just pausing, and when they’ve been interrupted mid-sentence. Anyone who has talked to a voice assistant knows the failure modes: the long awkward pause before it responds, the assistant that talks over you, the conversation that collapses the moment two people speak at once.
“We’re excited to welcome members of the Ultravox team and continue building the platform,” Inworld CEO Kylan Gibbs said in a statement. The acquisition lets Inworld adopt Ultravox’s technology directly into its character engine, which should make real-time voice chats noticeably less robotic.
A small team with an unusual resume
Ultravox had just 10 employees, according to LinkedIn — a reminder of how much leverage a small, senior team can have in this market. The founders’ backgrounds at Google and Apple gave them a head start on the hardest parts of speech: low-latency streaming, natural prosody, and the conversational dynamics that separate a voice interface people tolerate from one they enjoy.
The company’s most whimsical project hinted at the ambition. In December 2023, Ultravox launched HiSanta.ai, a website where users could chat with Santa Claus — speaking with a British accent — along with Rudolph, Mrs. Claus and other North Pole notables. The site is no longer active, but it demonstrated the core tech in the most demanding setting imaginable: a real-time character performance where latency and unnatural pauses would instantly break the illusion.
What happens to the team
It was unclear at announcement whether the entire team would join Inworld. Co-founder Zach Koch, most recently the company’s CEO, now lists himself as vice president of product at Shopify on LinkedIn — suggesting at least some of the team is scattering rather than moving to Mountain View en masse. That’s a common pattern in acqui-hires, and it puts pressure on Inworld to integrate the technology quickly, before the institutional knowledge disperses.
The bigger picture: voice is having a moment
The acquisition lands as voice AI moves from novelty to infrastructure — the same moment producing voice AI’s strangest success stories. Real-time speech models have crossed a quality threshold in the last two years, and the remaining problems are conversational, not acoustic: barge-in, backchanneling (“mm-hmm”), repair when the model mishears. Whoever solves the interaction layer owns the interface — whether that’s games, customer service, or the personal AI agents now raising billions to handle your phone calls. Inworld just bought a shortcut.
The state of voice AI in 2026
Voice interfaces have crossed a quality threshold that would have seemed like science fiction five years ago. Latency is down to hundreds of milliseconds, prosody — the rhythm and intonation of speech — is convincingly human, and real-time speech models can hold a conversation without the stilted turn-taking of early assistants. What’s left is the interaction layer: the subtle mechanics of human conversation that nobody notices until they’re wrong.
Consider what happens when you interrupt a person. They stop, yield, and pick up the thread — or they hold the floor with a raised finger and a “one second.” They backchannel constantly: the “mm-hmm” and “right” that signal attention without claiming a turn. They repair mishearings gracefully instead of restarting. Each of these behaviors is a research problem. Together, they’re the difference between talking to a machine and talking with something that feels present. Ultravox spent four years on exactly these problems — while on the memory side, open-source VoiceMem is giving voice agents a ‘left brain’ and a ‘right brain’.
What Inworld actually bought
Inworld’s existing technology already generates character dialogue, personality and memory. What it lacked was the conversational control plane — the real-time decisions about when to speak, when to yield and how to handle being interrupted. That’s Ultravox’s core IP: speech models purpose-built for live, full-duplex conversation rather than the half-duplex pattern of “you talk, then I talk” that most voice assistants still use.
The price of the technology reflects a broader market reality. Training a competitive speech model from scratch takes years and a specialized team; acquiring one takes a wire transfer. With ten employees and backing from top-tier firms, Ultravox was never going to out-distribute Inworld — but Inworld couldn’t replicate four years of speech research before its competitors shipped. In a market moving this fast, buying time is often the rational trade.
Why gaming is the proving ground
Games are where conversational AI faces its harshest test. A customer-service bot can pause, hedge and ask clarifying questions; a game character that does that breaks immersion instantly. Players interrupt, talk over dialogue, go silent mid-conversation and expect the character to react like a person would. If Ultravox’s turn-taking technology works in games, it works anywhere — which is presumably why Inworld, valued at $515 million on $119.8 million raised, decided it couldn’t wait to build it in-house.
The open question is integration speed. Acquisitions of tiny, senior teams often lose as much knowledge as they transfer, especially when founders scatter — and with Ultravox’s CEO now at Shopify, Inworld is racing the clock. The technology is only valuable if it ships inside Inworld’s character engine while the window is open. In voice AI in 2026, that window is measured in months.
Featured image: PhilyG (Phil McGowan) via Wikimedia Commons (Public domain)
