Inworld acquires Ultravox to add a voice-agent platform to its speech stack

Why it matters

Inworld has added the agent platform layer to its speech-model and inference business. Existing Ultravox customers get an immediate voice upgrade, while the acquisition gives Inworld a path to bring more of its stack into developers' applications.

Inworld AI acquired Ultravox, the platform for building real-time voice agents, bringing Ultravox team members into Inworld as it expands from speech models and infrastructure into the software developers use to build agents. Inworld announced the deal on X and detailed it in a company blog post dated September 30th, 2026.

Ultravox co-founder and former CEO Zach Koch said the two companies had already worked together on text-to-speech. Before Ultravox, Koch led product work at Shopify and Google, including teams working on Chrome and Pixel, according to GeekWire's 2023 report on Fixie, Ultravox's predecessor. In the acquisition announcement, Koch called Inworld the right home for the platform and said the Ultravox team would continue building it there.

The products cover different parts of a live voice interaction. Ultravox provides speech understanding, reasoning, task completion and controls for turn-taking and interruptions. Inworld supplies speech models, model serving and real-time inference. The intended combination puts the agent layer alongside the technology that listens and speaks, rather than leaving developers to assemble those pieces from separate providers.

Ultravox's own product description emphasizes direct speech processing without a traditional automatic speech recognition step. The company says that approach avoids converting speech to text before an AI model processes it, then converting the answer back to audio. Its platform also supports phone calls, web and native apps, and custom integrations through WebSockets and its native protocol. Those capabilities make Ultravox more than a voice interface: it is the developer-facing system for configuring and deploying agents that can take actions during a conversation.

The first integration is already live. Built-in Inworld voices on Ultravox now use Realtime TTS-2, while customers can keep their existing voice IDs and do not need to change code. Inworld says the upgrade adds optional controls for pace, tone and emotional delivery, with no additional charge. Customers using voices from other providers are unaffected by this change, according to the announcement.

That first step gives Inworld an immediate route for putting its voice model in a separate developer platform while keeping Ultravox available to existing users. Inworld has described its own broader product strategy as combining speech models with serving infrastructure and APIs; Ultravox adds an established agent-building surface to that stack. The deal's longer-term value will depend on whether the companies can extend the integration beyond the voice upgrade without disrupting developers who chose Ultravox as a standalone platform.

For developers evaluating the product today, Ultravox lists 30 free minutes and then usage-based pricing of $0.05 per minute on its pricing page. Its listed Pro plan is $100 per month and removes the hard concurrency cap for paid accounts. The TTS upgrade is separate from those existing platform charges: Inworld says it adds no new fee for Ultravox users with built-in Inworld voices.

The acquisition brings together companies with different funding histories. Inworld said in a company profile published in February 2026 that it had raised more than $125 million, with backers including Lightspeed, Kleiner Perkins, Founders Fund, CRV, Microsoft M12 and Meta. Fixie raised a $17 million seed round in 2023, led by Redpoint Ventures with Madrona Venture Group, Zetta Venture Partners, SignalFire, Bloomberg Beta and Kearny Jackson participating, according to GeekWire. The acquisition announcement gives no financial terms.

For Koch, the move carries forward a thesis Ultravox laid out in its company history: voice agents built by chaining speech recognition, text-based reasoning and speech generation can feel slow and awkward, so the company pursued a speech-native approach. Inworld's acquisition gives that work access to a company building its own speech models and inference systems. The near-term proof is narrow but concrete: existing Ultravox agents can use the updated Inworld voices without migration or an added charge. The broader bet is that developers will find a single integrated stack easier to ship and operate than a collection of separate voice and agent components.