Deepgram to acquire Daily, the company behind open-source voice agent framework Pipecat

Deepgram is acquiring Daily, the company behind Pipecat, the open-source voice agent framework, announced on 29 October at its inaugural Deepgram Speak event in San Francisco. Pipecat will stay open source and free under its existing BSD V2 license, and the companies are setting up an independent Pipecat Foundation. The release does not give a price.

What Deepgram gets from Daily

Deepgram says the deal brings together its speech models, Daily's global real-time network infrastructure and Pipecat's orchestration framework. Together they will underpin its new Voice Agent Platform, which the company describes as "the first complete, open, production-grade platform for conversational voice AI". Following the acquisition, the Daily team will join Deepgram.

Daily co-founder Kwindla Hultman Kramer becomes Deepgram's Head of Open Source. "Keeping Pipecat open was fundamental to our decision to join Deepgram," he said.

Deepgram CEO and co-founder Scott Stephenson said: "Acquiring Daily allows us to bring that expertise together to tackle the whole voice system and give developers a faster path from prototype to production."

Will Pipecat stay open and vendor neutral?

That is the commitment from both sides. Deepgram says it will invest in Pipecat's development and community while keeping it open source and vendor neutral, so developers can mix models, tools, network transport and APIs from different providers.

In Daily's own post, Kramer says the foundation will be run by a steering committee representing the whole real-time AI ecosystem, "including companies that compete with Deepgram". He adds that Daily's customers should see no change, with Deepgram selling, supporting and developing all of Daily's current products. Daily's post says Pipecat now has more than 200 integrations and almost 400 contributors.

A new open protocol for voice context

Deepgram also introduced the Real-Time Context Standard (RCS), an open protocol co-developed with the Pipecat team. The company says today's voice AI systems pass words from one model to the next but lose the audio context around them, such as tone, who is speaking and background sound.

RCS defines three types of metadata:

• context from the audio, such as a whisper or an unclear word
• instructions for how a language model should use those signals
• guidance for how the response should sound

Deepgram says that lets an agent ask for clarification when it did not hear clearly, or answer quietly to someone whispering. Adapters translate the signals for each model, so the same approach works across providers. Developers can test RCS and contribute to the specification through a demo repository on GitHub.

What else Deepgram announced at Speak

Chief Product Officer Ed Anuff was due to preview the forthcoming Deepgram Voice Agent Platform later the same day. The release says he would also introduce two new models built for live conversation, expanded language support and a Global API designed to improve connection reliability and reduce latency. Full details were to follow on Deepgram's blog.

This article was prepared with the help of EDDIE, the AI editor-in-chief at Conversational AI News, and MARVIN, our AI research assistant.