Deepgram Acquires Daily to Own the Voice Agent Stack

Deepgram said on Thursday it is acquiring Daily, the company behind Pipecat, one of the most widely used open-source frameworks for real-time voice AI. No price was disclosed. The commitment that matters is different: Pipecat stays open source, and it stays vendor-neutral.

What Deepgram actually bought

Pipecat is the plumbing of a voice agent. It moves audio and video through a pipeline that stitches speech-to-text, a language model, and text-to-speech together, and keeps latency low enough that a conversation still feels like a conversation. Daily, the company that maintains it, also sells hosted infrastructure for voice and video calls of any kind.

The two firms share history: both came out of the same Y Combinator batch in 2016. Daily had raised more than $62 million from Tiger Global, Nat Friedman, Kevin Weil and others, per PitchBook data cited by TechCrunch. Deepgram reached unicorn status earlier this year with a $130 million Series C.

The logic is straightforward. Deepgram sells the ears and the mouth, speech recognition and speech synthesis. Pipecat sells the nervous system that connects those parts to a model and a live call. Owning both means owning the whole path from microphone to answer, which is exactly where voice agents live or die.

The open-source promise

Daily co-founder Kwindla Kramer said the company had plenty of acquisition interest, but he chose Deepgram because CEO Scott Stephenson committed to keeping Pipecat open source "forever" and vendor-neutral, and because the engineering cultures matched. Stephenson framed openness as commercial strategy rather than charity: sit in the middle of a fast-growing ecosystem, be genuinely good at a few parts of it, and letting everything else stay open becomes an advantage.

Deepgram is also launching a Real-Time Context Standard, an open protocol for voice agents co-developed with the Pipecat team. It describes metadata carried alongside the audio, such as a whisper, instructions telling the model how to use those signals, and guidance on how the response should sound. In practice that is a spec for expressivity, the part of voice AI that pipelines usually flatten.

Why it matters for builders

If you run voice agents in production, the orchestration layer you chose just got a new owner. The stated vendor-neutrality is the thing to watch: it keeps Pipecat usable next to ElevenLabs voices, OpenAI or Anthropic models, and any telephony provider, but the roadmap now belongs to a speech vendor with its own models to sell.

The Real-Time Context Standard is the more interesting artifact. An open format for passing prosody and style signals through a pipeline is the kind of plumbing that stops being interesting precisely because everyone adopts it, and voice agents have needed a shared vocabulary for this since the first demos. For teams building phone agents, the practical question is no longer which framework supports the most models, but which one will carry that metadata through unchanged.