Skip to content
agentscamp
Guide · Voice

Realtime Voice Agents: Build on LiveKit, Buy Vapi, or Pipeline with Pipecat

The three ways to ship a realtime voice agent in 2026 — open infrastructure, managed platform, or OSS pipeline framework — and how speech-to-speech models change it.

2 min readAgentsCamp
Updated Jun 11, 2026
voice-agentsrealtimecomparisonwebrtc

Three postures cover realtime voice in 2026: build on LiveKit (open WebRTC infra + agents framework + telephony — maximum control), assemble with Pipecat (the OSS pipeline framework for custom STT→LLM→TTS flows), or buy Vapi (assistants live in an afternoon at a per-minute platform fee). Speech-to-speech realtime models slot into all three rather than replacing them.

Key takeaways

  • The decision is posture, not features: own the infrastructure (LiveKit), own the pipeline logic (Pipecat), or own neither and ship today (Vapi).
  • Latency budgets rule everything: a natural turn is well under a second total, split across STT, LLM time-to-first-token, and TTS time-to-first-audio — every architecture choice is a withdrawal from that budget.
  • Turn detection is the hidden boss: knowing when the user finished talking — LiveKit open-sourced a semantic model for it, Cartesia's Ink emits it natively, platforms bundle it.
  • Speech-to-speech models (realtime APIs) compress the pipeline but not the system — transport, telephony, tools, and interruptions still need a home, which is why they run INSIDE these stacks.
  • Economics flip with scale: per-minute platforms are unbeatable to start and expensive at volume; self-built stacks invert that curve.

Voice agents crossed the production threshold — a billion-plus calls on the major platforms — and the tooling sorted into three honest postures. The question isn't which is "best"; it's how much of the realtime stack you want to own.

The short list

PostureToolYou ownYou get
BuildLiveKitInfra + pipelineOpen source, max control, scale economics
AssemblePipecatPipeline logicOSS framework, provider freedom
BuyVapiConfigLive agents in an afternoon, per-minute fee

The three postures

Build on LiveKit when voice is core product. The Apache-2.0 WebRTC server plus the Agents framework covers transport, the STT→LLM→TTS pipeline or realtime speech-to-speech models, an open-sourced semantic turn-detection model, and Telephony 1.0 (SIP, transfers, scale) — with LiveKit Cloud as the managed escape hatch. The credential is hard to argue with: per LiveKit, ChatGPT's Voice Mode runs on this stack. Cost: real engineering; payoff: control and unit economics that improve with volume.

Assemble with Pipecat when the pipeline is your differentiation. The open-source Python framework composes voice flows from interchangeable pieces — any STT, any LLM, any TTS, custom logic between stages — without also adopting a media-server worldview. It pairs naturally with LiveKit or other transports underneath.

Buy Vapi when shipping beats owning. Assistant = prompt + model + voice + tools; attach a number; live. Turn-taking (vendor-claimed sub-600ms), interruptions, telephony, and multi-agent Squads come managed, at a platform fee per minute plus model costs (BYO keys pass through at cost). The 2026 traction — a $50M Series B, Amazon Ring routing all inbound calls through it — says the buy side is no toy. (Cartesia Line plays the same posture, vertically integrated on Cartesia's models.)

What actually decides quality

Whatever posture you pick, the same three system properties make or break the agent. The latency budget: a natural conversational turn is well under a second — STT endpointing + LLM time-to-first-token + TTS time-to-first-audio, measured p95, decides whether the agent feels human or like hold music. Turn detection: knowing when the user finished (versus paused) is the hardest perceptual problem in the stack — LiveKit's open semantic model, Ink's native turn events, and platform bundles are all answers to it. Interruption handling: users barge in; the agent must stop talking, cheaply discard in-flight generation, and listen — a transport-and-state problem no model solves alone.

Start by posture (own infra / own pipeline / own nothing), prototype on the buy side if speed matters, and revisit the build math when minutes get expensive. The component-level walkthrough — models, prompts, and the pipeline's failure modes — is How to Build a Voice Agent, and the voice-agent-engineer agent owns exactly this build.

Frequently asked questions

Should I use a speech-to-speech model or the STT→LLM→TTS pipeline?
Increasingly both exist in the same system. Realtime speech-to-speech models give the most natural turn-taking and lowest conversational latency; the classic pipeline gives model choice per stage, easier tool-calling control, and cheaper economics. The frameworks treat it as a config choice — LiveKit Agents and Pipecat run either; Vapi abstracts it — so the architecture decision (build/assemble/buy) matters more than the model topology.
What does a production voice agent actually consist of?
Transport (WebRTC or telephony), speech recognition with endpointing, turn detection, the agent brain (LLM + tools + memory), speech synthesis, and interruption handling — plus observability over all of it. The model calls are the easy 30%; the realtime systems engineering is why platforms and frameworks exist.
What does Cartesia Line change?
It's a fourth posture: vertically integrated. Line runs Cartesia's own Sonic TTS and Ink STT in a hosted agent platform with per-minute pricing — competing with Vapi on convenience while owning the models. Compelling if Cartesia's latency thesis is your priority; the trade is provider flexibility.

Related