Deepgram acquires Daily, maker of the open-source Pipecat voice agent framework

Deepgram announced on 2 November that it is acquiring Daily, the company behind Pipecat, an open-source framework for building voice agents, and behind a real-time network that carries audio between users and agents. No price, structure or financial terms were disclosed. The combination pairs Deepgram's speech models with Pipecat's orchestration layer and Daily's transport, and will form the basis of a planned Deepgram Voice Agent Platform that the company describes as open and model-agnostic, letting developers mix models, tools, transport and APIs from different providers. Pipecat stays open source under its existing BSD licence, and the companies say they intend to set up an independent Pipecat Foundation. The Daily team joins Deepgram, with co-founder Kwindla Hultman Kramer becoming Head of Open Source. Alongside the deal, Deepgram introduced a Real-Time Context Standard, an open protocol co-developed with the Pipecat team for passing audio context such as tone or whispering between models, and previewed two live-conversation models, expanded language support and a Global API aimed at connection reliability and latency.

Why it matters for digital humans: Almost every real-time digital human is a voice agent with a face attached, and Pipecat is one of the pieces doing the attaching — the layer that holds speech recognition, the model, text-to-speech and often the avatar renderer in one pipeline. Consolidating that orchestration layer, the transport and the speech models under a single vendor is convenient, and it also narrows the number of independent parties in a stack many avatar products quietly depend on. The announced Pipecat Foundation is the thing to watch: whether governance actually moves out of Deepgram will determine how portable those pipelines stay. The context-passing protocol is the more immediately interesting piece for avatar builders, since paralinguistic signals such as tone are exactly what a face has to render and exactly what a text-in-the-middle pipeline discards.