Deepgram moves to acquire Daily as Pipecat gains a voice-context standard

Deepgram announced on 29 October that it is acquiring Daily, the company behind the Pipecat voice-agent framework. It also outlined plans for an independent Pipecat Foundation and the Real-time Context Standard, or RCS, alongside its move towards a broader voice-agent platform.[1]

The acquisition brings three parts of a conversation under one commercial roof: Deepgram’s speech models, Daily’s network for live audio and video, and Pipecat’s software for connecting models and handling conversations. Deepgram describes its integrated Voice Agent Platform as forthcoming.[1]

Keeping a choice of models

Daily co-founder Kwindla Hultman Kramer says Deepgram will continue selling, supporting and developing Daily’s products. The proposed foundation will have a steering committee drawn from the wider real-time AI ecosystem, including Deepgram’s competitors. This makes governance consequential: the framework connects services that may compete with its new commercial owner.[2]

Pipecat’s repository currently uses the BSD-2-Clause licence. Deepgram has committed to keeping the framework free, open source and available for mixing models, tools and network providers.[5][1]

What RCS adds to a conversation

A conventional voice pipeline turns speech into text, generates an answer, then turns that answer back into speech. RCS adds information about how a person sounded and how the reply should be delivered to the text passed between these components. Version 0.1 is an alpha whose formats may change before 1.0.[3]

Support is selective: a component can comply without implementing every context type. Perception producers also define their own signal names. Developers therefore still need to match the capabilities of connected models; the shared format alone does not make every signal interchangeable. Speech consumers must strip unsupported or interrupted tags instead of reading them aloud.[3]

Inside the published demonstration

The demo’s sample configuration connects Deepgram transcription, Pipecat Tone, a language model and ElevenLabs or Cartesia speech generation. Its example shop handles order enquiries using three orders held in memory. Developers can inspect that integration in the repository; it is a demonstration, not a reported customer deployment.[4]

The repository also draws a useful testing boundary: its scripted evaluations run in text mode and check tool calls and service rules. They do not evaluate the spoken interaction. Teams can inspect how the pieces connect today, then assess their own audio pipeline separately.[4]

Report an error or suggest a correction