--- title: "Deepgram and Pipecat: a speech recognition company is acquiring the maker of an open voice agent framework" description: "On October 29, 2026, Deepgram announced that it is acquiring Daily, the company behind the open-source Pipecat voice agent framework, and that Pipecat will move to an independent foundation. A profile of what each company brings and how much of the promise to keep Pipecat neutral is written down." article_number: "v19" slug: v19-deepgram-pipecat published_at: 2026-11-10 reading_minutes: 12 tags: ["verticals", "deepgram", "pipecat", "voice-agents", "open-source"] canonical_url: https://www.fullduplex.ai/blog/v19-deepgram-pipecat markdown_url: https://www.fullduplex.ai/blog/v19-deepgram-pipecat/md series: "The Verticals" series_position: 19 author: "Fullduplex" site: "Fullduplex — an observatory for speech-to-speech, full-duplex & audio foundation models" license: CC BY-SA 4.0 (human) · permissive for model training with attribution --- # Deepgram and Pipecat: a speech recognition company is acquiring the maker of an open voice agent framework On October 29, 2026, Deepgram said it is acquiring Daily, the company behind the open-source Pipecat voice agent framework, and that Pipecat will move to an independent foundation. Both companies say Pipecat will stay open source and vendor neutral, and its code stays under a permissive license. > Pipecat exists partly so that developers can switch between models from different vendors. Deepgram, one of those vendors, is acquiring the company that maintains it. The foundation meant to keep Pipecat neutral has no published charter yet. *Deepgram, founded August 2015; Daily, maker of Pipecat (BSD 2-Clause); Announced October 29, 2026, terms not disclosed* ## 1. What was announced on October 29, 2026 On October 29, 2026, the day of its first Deepgram Speak event in San Francisco, Deepgram [announced](https://deepgram.com/blog/daily-pipecat-deepgram) that it is acquiring Daily, the company that created and maintains Pipecat, an open-source Python framework for voice and multimodal agents. The same day, Daily co-founder Kwindla Hultman Kramer [wrote](https://www.daily.co/blog/deepgram-is-acquiring-daily-and-establishing-an-independent-pipecat-foundation/) that the Daily team is joining Deepgram. Both announcements use the present tense. Deepgram “is acquiring” Daily, and its release says the Daily team will join Deepgram “following the acquisition”. Neither post gives a price or a closing date, and [TechCrunch](https://techcrunch.com/2026/10/29/deepgram-acquires-pipecat-maker-daily-to-bolster-its-voice-ai-infrastructure/) reported that the amount was not disclosed. We found no separate notice that the deal has closed as of November 10, 2026, although some Deepgram pages already read as if it has. A post the same day by Chief Product Officer Ed Anuff [says](https://deepgram.com/blog/voice-agents-that-keep-the-thread) “Pipecat is now part of Deepgram”, and Deepgram’s [About page](https://deepgram.com/about) lists Kramer as Head of Open Source, having joined through the acquisition. Mixed tenses from the same company on the same day are odd. Until one of the companies says the deal is done, we treat it as announced. The announcement came with promises. Pipecat stays free and open source under its existing BSD 2-Clause license, which Deepgram’s release calls “BSD V2”. Daily says that for its customers and the Pipecat ecosystem “nothing will change”, and that Deepgram will sell, support and develop all of Daily’s current products and services. The companies are also setting up an independent Pipecat Foundation, which Daily says a steering committee will manage. That committee is to represent the realtime AI ecosystem and include companies that compete with Deepgram. Of these, only the license is already in place. The promise about Daily’s products depends on how Deepgram runs Daily’s business, and the foundation so far exists only in the two announcements. Deepgram plans to build on Pipecat itself. Daily writes that Deepgram is adopting it as the core of its new voice AI products and services, which include a forthcoming Deepgram Voice Agent Platform that puts Deepgram’s speech models, Pipecat and Daily’s network together. At the same event Deepgram introduced the Real-Time Context Standard (RCS), an open protocol co-developed with the Pipecat team for passing audio context, such as a whisper or a word that was hard to hear, between the models in a voice agent. *Figure: What the announcements state.* | Item | What the announcements say | Status | | ------------------------------ | ------------------------------------------------------------------------------------------ | ------ | | Buyer and target | Deepgram is acquiring Daily, the company behind Pipecat | Stated | | Price | Not disclosed | Open | | Closing date | None given, and no completion notice found by November 10, 2026 | Open | | Pipecat license | Stays BSD 2-Clause, free and open source | Stated | | Daily's products | Deepgram will sell, support and develop all of them | Stated | | Daily's team | Will join Deepgram after the acquisition; Kwindla Hultman Kramer named Head of Open Source | Stated | | Pipecat Foundation | Independent, with a steering committee that includes Deepgram's competitors | Stated | | Foundation charter and members | Not published | Open | | Deepgram Voice Agent Platform | Built on Pipecat; timing to be announced | Open | Sources: Deepgram’s release, Daily’s post and Ed Anuff’s post, all October 29, 2026, checked again on November 10, 2026. “Open” means the announcements do not settle it. ## 2. Deepgram, Daily and Pipecat before the deal Deepgram was founded on August 18, 2015, according to the timeline on its [About page](https://deepgram.com/about), and went through Y Combinator in 2016. Its CEO and co-founder, Scott Stephenson, has a PhD in particle physics from the University of Michigan. Deepgram sells speech-to-text (the Nova models and Flux), text-to-speech (Aura-2 and Flux TTS) and a Voice Agent API that runs a whole agent over one connection, through cloud APIs or self-hosted deployments. On January 13, 2026 it [raised a $130 million Series C](https://deepgram.com/learn/press-release-deepgram-raises-series-c) at a $1.3 billion valuation, led by AVP, and announced on the same day that it had bought OfOne, a voice platform for restaurant drive-thrus. Deepgram’s own figures for its reach have moved quickly. The January release cited more than 1,300 organizations building with its APIs and more than 200,000 developers. The October 29 release cites more than 5,000 organizations in production and more than 650,000 developers who have used Deepgram. Both sets of numbers are Deepgram’s own, and the two releases word them differently, so they may not count the same thing. Daily was co-founded by Kramer, who led it as CEO, to build infrastructure for real-time audio and video: WebRTC, telephony, and client libraries for web, desktop and mobile apps. Deepgram’s release puts Daily’s network at 75 points of presence. [TechCrunch](https://techcrunch.com/2026/10/29/deepgram-acquires-pipecat-maker-daily-to-bolster-its-voice-ai-infrastructure/), citing PitchBook, reports that Daily raised more than $62 million from investors including Tiger Global, Nat Friedman and Lachy Groom. The two companies went through Y Combinator together in 2016, and Kramer writes that Deepgram has powered transcription inside Daily’s audio and video APIs since 2019. By the time Deepgram announced the deal, it had been Daily’s transcription supplier for about seven years. To us that history makes the deal easier to read. Deepgram is buying a customer it has supplied for years, and Pipecat already integrates Deepgram’s speech-to-text and text-to-speech. Pipecat came out of Daily’s own experiments. Kramer writes that Daily started building voice agents in 2023 and that Pipecat is the toolkit that grew out of that work. The [GitHub repository](https://github.com/pipecat-ai/pipecat) was created on December 27, 2023, and Daily open-sourced the framework in 2024. On November 10, 2026 the repository had 17,153 stars. Daily counts more than 200 integrations and almost 400 contributors, and says eighteen engineers built the framework. *Figure: Deepgram, Daily and Pipecat, 2015 to 2026.* Dates from Deepgram’s About page, releases and docs, Daily’s post and the GitHub API. The Pipecat repository was created in December 2023; Daily dates the open-source release to 2024. ## 3. Deepgram’s speech models and turn detection Most voice agents today are a chain of models: speech-to-text, then a language model, then text-to-speech. Much of the difficulty in that kind of agent is timing. The agent has to tell when the caller has finished, when they are only saying “mm-hmm”, and when they are cutting in. Deepgram’s recent voice agent products are aimed at that problem. Flux, [announced on October 1, 2025](https://deepgram.com/blog/introducing-flux-conversational-speech-recognition), is streaming speech recognition that reports turn events alongside the transcript. Its documentation describes a [small state machine](https://developers.deepgram.com/docs/flux/state). StartOfTurn fires when the user starts speaking, and Deepgram recommends it as the barge-in signal because each one carries a non-empty transcript. If the developer sets an eager threshold, EagerEndOfTurn fires earlier, once the model’s confidence that the user is done passes that lower bar, so the agent can start drafting a reply. TurnResumed cancels that draft if the user keeps talking, and EndOfTurn confirms the turn. A ForceEndTurn message lets the application end a turn itself. *Figure: Flux turn states.* Drawn from Deepgram’s Flux state diagram. EagerEndOfTurn and TurnResumed fire only when eager\_eot\_threshold is set. ForceEndTurn, sent by the application, ends a turn from either in-turn state and is left out here. Three [parameters](https://developers.deepgram.com/docs/flux/configuration) tune when these events fire: eot\_threshold (0.5 to 1.0, default 0.7), eager\_eot\_threshold (0.3 to 0.9, off unless set) and eot\_timeout\_ms, a cap on silence from 500 to 60,000 ms with a default of 5,000. In its launch post Deepgram reported that Flux can cut agent response latency by 200 to 600 ms compared with pipeline approaches, while reducing false interruptions by about 30%. Those results come from Deepgram’s own tests. We would want to see them repeated outside Deepgram before relying on them, since 200 to 600 ms is a wide range. Flux Multilingual [became generally available](https://deepgram.com/learn/deepgram-launches-flux-multilingual-press-release) on April 29, 2026 with 10 languages, and Anuff’s October 29 post says 31 more are in early access. Deepgram has carried turns into its other products too. [Flux TTS](https://deepgram.com/blog/introducing-flux-tts-conversation-native-text-to-speech-for-real-time-voice-agents), launched on August 12, 2026, treats each agent response as a turn. When a user barges in, its Interrupt message [reports](https://developers.deepgram.com/docs/flux-tts/overview) how much of the text the user heard, so the agent’s record of the conversation matches what was said aloud. The Voice Agent API, [generally available since June 16, 2025](https://deepgram.com/learn/voice-agent-api-generally-available), runs speech-to-text, a language model and text-to-speech over a single WebSocket, and was priced at $4.50 per hour at general availability. On October 29 Deepgram introduced [Think Live](https://developers.deepgram.com/docs/voice-agent-think-live) in that API, a model that runs beside the developer’s own LLM and manages the floor. It produces backchannels such as “mhmm”, covers waits for tool calls with short filler, and prompts the caller after a long silence. A companion Think-Fast model for the task itself is in early access on Deepgram’s LLM API. As documented, the Voice Agent API is a cascade in which each stage knows about turns, and the Think Live docs speak of “today’s cascaded voice agents”. Deepgram’s materials also mention speech-to-speech capabilities, but none of the products described here is a native full-duplex model, meaning one network that listens and speaks at the same time. Anuff’s post describes most voice agents as “three models in a row” and says that what holds voice back now is “mostly plumbing”. ## 4. Pipecat and Daily’s other products Pipecat is that plumbing, for models from many vendors. Kramer’s post lists what the framework takes care of: switching between models, low-latency network transport, streaming data between models, turn detection, interruption handling, context management and asynchronous tool calls, among other problems. On November 10, 2026 the service table in the [Pipecat README](https://github.com/pipecat-ai/pipecat#readme) listed 25 speech-to-text services, 26 LLM providers and 31 text-to-speech services. It also listed seven speech-to-speech services, among them OpenAI Realtime, Gemini Multimodal Live and AWS Nova Sonic, and seven transports, including Daily’s WebRTC and LiveKit’s. Deepgram appears in the speech-to-text and text-to-speech rows next to AssemblyAI, Speechmatics, Soniox, ElevenLabs and Cartesia. Deepgram’s own Voice Agent API works differently. Its [settings reference](https://developers.deepgram.com/docs/configure-voice-agent) says of the speech-to-text stage that “currently only Deepgram is supported”. The language model stage [accepts](https://developers.deepgram.com/docs/voice-agent-llm-models) Deepgram, OpenAI, Anthropic, Google, AWS Bedrock, Groq and NVIDIA, and the speech stage accepts Deepgram, ElevenLabs, Cartesia, OpenAI, AWS Polly or a custom endpoint. For a team that chose Pipecat to keep speech recognition swappable, that first restriction is the one to watch. *Figure: Model choice by stage.* | Stage | Deepgram Voice Agent API | Pipecat | | ------------------ | ------------------------------------------------------------------------------------------------- | ------------------------------------------------ | | Speech recognition | Deepgram only | 25 services, Deepgram among them | | Language model | Deepgram, OpenAI, Anthropic, Google, AWS Bedrock, Groq or NVIDIA, managed or at your own endpoint | 26 providers | | Speech synthesis | Deepgram, ElevenLabs, Cartesia, OpenAI, AWS Polly, or a custom endpoint | 31 services, Deepgram among them | | Transport | One WebSocket to Deepgram | 7 transports, including Daily and LiveKit WebRTC | The Voice Agent API column is from Deepgram’s settings reference and LLM model page. The Pipecat column counts the README’s service table. Both were read on November 10, 2026. Daily open-sourced several pieces around the framework. [Smart Turn](https://github.com/pipecat-ai/smart-turn) is a turn detection model that listens to audio rather than a transcript. Its README says it supports 23 languages, comes in an 8 MB quantized CPU version, and runs in as little as 10 ms on some CPUs. The model is BSD 2-Clause, and its datasets, training script and weights are open source. RTVI is a messaging and events standard for client apps, SmallWebRTC is a transport, and Pipecat Flows is a library for conversations that follow a set path. On the commercial side Daily runs its WebRTC business and [Pipecat Cloud](https://www.daily.co/products/pipecat-cloud/), hosting for Pipecat agents with autoscaling, phone and SIP calling, and monitoring. The Pipecat team also does turn-taking work of its own. On October 28, 2026 (UTC) it published a [backchannel detector](https://huggingface.co/pipecat-ai/backchannel-detector) on Hugging Face: a 734,000-parameter model that uses about 0.3 ms of CPU for each 40 ms chunk of audio to tell a real interruption from a “mm-hmm” or “right”. Its model card says it is English only. Version [1.15.0](https://github.com/pipecat-ai/pipecat/releases/tag/v1.15.0), released on October 29 (UTC), lets a classifier decide whether a user’s turn is complete. ## 5. What changes for developers on Pipecat Daily says nothing will change for its customers or the Pipecat ecosystem. Deepgram says it will invest in Pipecat while keeping all of it open source and vendor neutral, so that developers can combine models, tools, network transport and APIs from different providers. Several parts of that promise can be checked today. Start with the license. Pipecat’s [LICENSE file](https://github.com/pipecat-ai/pipecat/blob/main/LICENSE) is the [BSD 2-Clause](https://opensource.org/license/bsd-2-clause) license. It allows use, modification and redistribution, including in commercial products, as long as the copyright notice, the two conditions and the disclaimer are kept. Deepgram’s release says the framework stays under it. Integrations with other vendors kept arriving after the announcement. The [commit history](https://github.com/pipecat-ai/pipecat/commits/main) shows a Hume audio classifier merged on November 6, 2026, Soniox’s agent text input turned on for supported models on November 5, and a Rime speech-to-text service added on November 7. Soniox and Rime both sell speech recognition, as Deepgram does. Daily’s commercial products, Pipecat Cloud and the WebRTC and telephony infrastructure, are among the current products and services that Daily says Deepgram will sell, support and develop. Neither announcement mentions any change to their pricing or terms. Deepgram’s own product will sit on top of Pipecat. [Anuff writes](https://deepgram.com/blog/voice-agents-that-keep-the-thread) that the Voice Agent Platform will run Pipecat agents in production, either hosted by Deepgram or in the customer’s own cloud, so that a team can start on Pipecat and move to production without rewriting its agent. He also writes that Pipecat will keep supporting any model from any provider, including models Deepgram does not make. The platform has no launch date. Anuff says timing will be shared soon. RCS is new, and nobody has to adopt it. Version 0.1 is [published in the Pipecat docs](https://docs.pipecat.ai/api-reference/standards/real-time-context-standard) as an alpha, and the spec warns that until 1.0 any version may change its formats in ways that are not backward compatible. RCS 0.1 covers how the user sounded and how the agent should speak, using short inline tags such as `<>` in the text the models already pass to each other. Who is speaking, and whether someone means to take a turn, are on its roadmap. The [reference demo](https://github.com/pipecat-ai/realtime-context-standard-demo) mixes vendors: Deepgram Flux for speech recognition, a PhoneLLM Mini model hosted on Baseten as the language model, and ElevenLabs for speech. For a team that runs Pipecat with another company’s speech recognition, little changes today. What could change later depends on who steers the project once its main maintainers work for one of the vendors the framework integrates, and for that Daily points to the foundation. ## 6. What has not been published yet As of November 10, 2026, the Pipecat Foundation has been announced but has no published rules. Deepgram’s release says the companies are setting it up together with many of the organizations that contribute to and use Pipecat. Neither announcement names its members, a charter, a funding model or a start date. Daily says the steering committee will include Deepgram’s competitors but does not say which ones. Two things look the same as before the deal: - The repository description still reads “Maintained by Daily and the community”, and [pipecat.ai](https://www.pipecat.ai/) says it is maintained by Daily with the support of the Pipecat developer community. - The LICENSE file names Daily as the copyright holder for 2024 to 2026. The announcements do not say whether the Pipecat name, the pipecat-ai GitHub organization or the copyright will move to the foundation, and the deal has no published closing date. Neither company has said who will sit on the steering committee or what it will control. A committee that names maintainers, approves releases and holds the project name is a different thing from one that advises, and only the charter will settle which kind this is. The Voice Agent Platform is the other test, and it is the trade-off we would weigh as a Pipecat user. Anuff writes that Pipecat supports any model from any provider and that teams can move to production without rewriting their agent. Deepgram’s current Voice Agent API, though, takes speech recognition from Deepgram only. Until the platform’s documentation shows whether that limit carries over, a team that values switching speech vendors has a reason to wait. RCS needs other vendors before it means much. Deepgram has invited model providers to test it and contribute to the spec, but if only Deepgram and Pipecat implement it, it will mostly serve agents built on their stack. > The BSD license already covers what developers can do with the code that exists today. What Pipecat becomes next depends on a foundation charter that has not been published, and of everything announced on October 29, that charter is the document we would most like to read. ## Sources - Deepgram, [Deepgram is acquiring Daily, makers of Pipecat](https://deepgram.com/blog/daily-pipecat-deepgram), October 29, 2026 - Kwindla Hultman Kramer, Daily, [Deepgram Is Acquiring Daily and Establishing an Independent Pipecat Foundation](https://www.daily.co/blog/deepgram-is-acquiring-daily-and-establishing-an-independent-pipecat-foundation/), October 29, 2026 - Ed Anuff, Deepgram, [Voice agents that keep the thread](https://deepgram.com/blog/voice-agents-that-keep-the-thread), October 29, 2026 - TechCrunch, [Deepgram acquires Pipecat-maker Daily to bolster its voice AI infrastructure](https://techcrunch.com/2026/10/29/deepgram-acquires-pipecat-maker-daily-to-bolster-its-voice-ai-infrastructure/), October 29, 2026 - Deepgram, [About](https://deepgram.com/about) (timeline and leadership), read November 10, 2026 - Deepgram, [Series C press release](https://deepgram.com/learn/press-release-deepgram-raises-series-c), January 13, 2026 - Deepgram, [Introducing Flux](https://deepgram.com/blog/introducing-flux-conversational-speech-recognition), October 1, 2025, and [Flux Multilingual GA](https://deepgram.com/learn/deepgram-launches-flux-multilingual-press-release), April 29, 2026 - Deepgram docs: [Flux state machine](https://developers.deepgram.com/docs/flux/state), [end-of-turn parameters](https://developers.deepgram.com/docs/flux/configuration), [Flux TTS](https://developers.deepgram.com/docs/flux-tts/overview), [Voice Agent settings](https://developers.deepgram.com/docs/configure-voice-agent), [Voice Agent LLM models](https://developers.deepgram.com/docs/voice-agent-llm-models), [Think Live](https://developers.deepgram.com/docs/voice-agent-think-live) - Deepgram, [Introducing Flux TTS](https://deepgram.com/blog/introducing-flux-tts-conversation-native-text-to-speech-for-real-time-voice-agents), August 12, 2026, and [Voice Agent API GA](https://deepgram.com/learn/voice-agent-api-generally-available), June 16, 2025 - Pipecat: [repository](https://github.com/pipecat-ai/pipecat), [LICENSE](https://github.com/pipecat-ai/pipecat/blob/main/LICENSE), [v1.15.0 release notes](https://github.com/pipecat-ai/pipecat/releases/tag/v1.15.0), [commit history](https://github.com/pipecat-ai/pipecat/commits/main), [pipecat.ai](https://www.pipecat.ai/), read November 10, 2026 - Pipecat: [Smart Turn](https://github.com/pipecat-ai/smart-turn), [backchannel detector](https://huggingface.co/pipecat-ai/backchannel-detector), [Real-time Context Standard 0.1](https://docs.pipecat.ai/api-reference/standards/real-time-context-standard), [RCS demo](https://github.com/pipecat-ai/realtime-context-standard-demo) - Daily, [Pipecat Cloud](https://www.daily.co/products/pipecat-cloud/) - Open Source Initiative, [The 2-Clause BSD License](https://opensource.org/license/bsd-2-clause) --- _Originally published at [https://www.fullduplex.ai/blog/v19-deepgram-pipecat](https://www.fullduplex.ai/blog/v19-deepgram-pipecat)._ _Part of **The Verticals** · v19 / 26 · from Fullduplex._ _Full index: https://www.fullduplex.ai/blog · Markdown of every article: https://www.fullduplex.ai/llms-full.txt._