Deepgram to buy Daily, the company behind Pipecat
Deepgram is acquiring Daily, the company that created the open-source Pipecat framework. The announcement came at Deepgram Speak, the company's first event of its kind, held in San Francisco.
The deal puts speech models, real-time communications infrastructure and voice-agent orchestration under one roof. Daily brings a global network of 75 points of presence. Pipecat connects the models and services that make up a voice agent and handles streaming between them, turn detection, interruptions, context management and asynchronous tool calls. Daily co-founder Kwindla Hultman Kramer becomes Head of Open Source at Deepgram, and the rest of the Daily team joins as well.
Deepgram says Pipecat stays free, open source and vendor neutral. It remains under the BSD 2-Clause license. An independent Pipecat Foundation is being set up, with a steering committee meant to represent the wider ecosystem, competitors of Deepgram included. Daily says Deepgram will keep selling, supporting and developing its existing products, among them WebRTC infrastructure, telephony and cross-platform libraries.
The two companies already know each other. Both went through Y Combinator in 2016, and Deepgram has supplied transcription for Daily's audio and video APIs since 2019.
A standard for how speech sounds
A typical pipeline passes a transcript from speech recognition to a language model, then hands the generated text to a speech synthesizer. The words survive, but information about how they were spoken can drop out. Deepgram and the Pipecat team created the Real-Time Context Standard, or RCS, to carry those signals through the pipeline.
Perception information travels with the user's text. Expression tags tell the agent how to speak, covering delivery, emphasis, pauses and vocal sounds. Examples include recognizing an unclear word and asking for clarification, or adjusting an agent's delivery when a user speaks quietly. The format is built for streaming, so speech generation can run sentence by sentence and unfinished markup can be dropped if an interruption cuts a response short.
The published RCS 0.1 specification is an alpha. It covers how the user sounded and how the agent should respond vocally. Speaker identification and signals about whether someone intends to take a turn are still on the roadmap.
Deepgram also announced two language models. Think-Live decides when an agent should listen, wait or respond, and is generally available through the Voice Agent API. Think-Fast handles business rules, tool calls and replies meant to be spoken, and is in early access through the Deepgram LLM API.
The company added multilingual coverage of 55 languages for Nova-3 Multilingual, plus more Flux STT and TTS languages. The Voice Agent Platform is still forthcoming, with deployment planned on Deepgram hosting or in customers' own cloud environments. A Global API, aimed at more reliable connections and lower latency, is also on the way.
Source: audioXpress