Which cloud providers handle real-time audio streaming to LLM apps under 300ms latency?

As of October 2026 the four main choices are LiveKit Cloud (WebRTC plus hosted agents, $0.01 per agent-session minute after the plan allowance), Daily's Pipecat Cloud ($0.01 per active minute on agent-1x, 1:1 WebRTC voice transport free), Twilio ConversationRelay ($0.07 per minute plus normal Twilio voice minutes, for phone calls) and Cloudflare Realtime (SFU and TURN at $0.05 per GB of egress, first 1,000 GB a month free). The transport adds only tens of milliseconds when the user is near an edge region; the 300 ms budget is mostly spent by speech recognition, the model's first token and speech synthesis, so pair the transport with a speech-to-speech model or streaming STT and TTS.

Is WebRTC or WebSockets better for streaming audio to an LLM?

WebRTC for anything a person talks to live. It runs over UDP with jitter buffers, echo cancellation and packet-loss concealment, so a lost packet does not stall the stream the way it does on a TCP WebSocket. WebSockets are fine server to server, for example between your backend and a model API, and for telephony bridges such as Twilio Media Streams.

What does a voice AI minute cost on LiveKit or Pipecat Cloud?

The platform part is about one cent a minute on both: LiveKit Cloud charges $0.01 per agent-session minute beyond the 1,000 (Build), 5,000 (Ship, $50 a month) or 50,000 (Scale, $500 a month) included minutes, and Pipecat Cloud charges $0.01 per active minute for an agent-1x container. Speech recognition, the LLM and speech synthesis are billed on top, which LiveKit's own calculator puts at roughly $0.02 to $0.04 a minute for a typical cascaded stack.

Can I get under 300 ms end to end?

Only with care. Keep the agent in the same region as the user and the model, use a speech-to-speech model or streaming STT and TTS, and measure time from end of user speech to first audio byte. Cascaded STT to LLM to TTS pipelines commonly land above 300 ms; native audio models such as GPT-Live-1, Gemini 3.8 Live and Grok Voice 3.0 remove two hops.

Quick Answer

Cloud Providers for Real-Time Audio Streaming to LLM Apps

Published: October 20, 2026

The short answer

For a voice or audio app that talks to an LLM, pick the transport by where the audio comes from, then pick the model by latency. As of October 20, 2026:

LiveKit Cloud — the default for web and mobile voice agents: WebRTC rooms, hosted agent workers, SIP telephony and LiveKit Inference (STT, LLM, TTS behind one key) in one bill. $0.01 per agent-session minute after the plan allowance.
Pipecat Cloud (Daily) — the best fit if you already write agents in the open-source Pipecat framework: $0.01 per active minute (agent-1x), Daily WebRTC voice transport free for 1:1 sessions, PSTN dial-in $0.018 a minute.
Twilio ConversationRelay — the shortest path for phone calls: Twilio handles the call, speech recognition and synthesis and streams text to your LLM over a WebSocket. $0.07 a minute plus standard Twilio voice minutes.
Cloudflare Realtime — the cheapest raw transport at scale: an anycast SFU and TURN at $0.05 per GB of egress with 1,000 GB a month free, plus a WebSocket adapter to bridge audio to your backend. You build the agent layer yourself.

The transport is rarely what blows a 300 ms budget. Speech recognition, the model’s time to first token and speech synthesis are. This page covers the transport and hosting layer; for the models themselves see Grok Voice 3.0 vs GPT-Live-1 vs Gemini 3.8 Live, and for fast text inference see best AI infrastructure providers for real-time inference.

The comparison

Prices are from each vendor’s pricing page on October 20, 2026.

Provider	Audio enters via	Agent hosting	Platform price	Model costs	Best for
LiveKit Cloud	WebRTC, SIP, phone numbers	Yes (Python/Node agents, 1-second metering, 10-second minimum)	Build $0 (1,000 agent min) · Ship $50/mo (5,000) · Scale $500/mo (50,000) · then $0.01/min	LiveKit Inference credits ($2.50 / $5 / $50 included) or bring your own keys	Web and mobile voice agents, one vendor for media + agents
Pipecat Cloud (Daily)	Daily WebRTC, SIP $0.003–$0.02/min, PSTN $0.018/min, Twilio, WebSocket	Yes (containers: agent-1x/2x/3x)	$0.01 / $0.02 / $0.03 per active min; reserved warm instances $0.0005/min (1x)	Bring your own (80+ STT, LLM, TTS providers)	Teams on the open-source Pipecat framework
Twilio ConversationRelay	PSTN and SIP calls	No — your LLM app runs elsewhere and receives text over WebSocket	$0.07/min + Twilio voice minutes	Your LLM; STT/TTS handled by the relay	Phone agents on an existing Twilio estate
Cloudflare Realtime	WebRTC SFU, TURN, WebSocket adapter	No (Workers, Durable Objects and Containers billed separately)	$0.05/GB egress; first 1,000 GB/mo free; ingress free	Workers AI or any API	High-volume, build-it-yourself audio transport

Where the 300 milliseconds go

Measure from the moment the user stops speaking to the first byte of the agent’s audio. In a cascaded pipeline, that interval is the sum of:

End-of-turn detection — deciding the user has finished. LiveKit includes its turn-detection model free for agents hosted on its cloud; Pipecat Cloud includes Krisp VIVA noise and voice isolation free up to 10,000 active minutes a month.
Final speech-to-text result — streaming STT returns partial transcripts as the user talks, so only the last fragment is on the critical path.
LLM time to first token — usually the largest single item. Small, fast models and prompt caching matter more than the transport.
Text-to-speech time to first audio — streaming TTS starts speaking before the sentence is complete.
Network — two legs, user to edge and edge to model. Keep the agent worker, the model endpoint and the user in the same region.

A speech-to-speech model collapses steps 2–4 into one: GPT-Live-1 is $0.05 a minute, Grok Voice 3.0 $0.08 a minute, and Gemini 3.8 Live bills audio at about $0.005 a minute in and $0.018 out (prices from our voice API price table, verified October 17, 2026). All four providers above can carry either architecture.

Picks by job

A voice assistant in a web or mobile app: LiveKit Cloud. Media, agent hosting, telephony and observability in one place; the free Build plan covers 1,000 agent minutes to prototype.
You already have a Pipecat agent: Pipecat Cloud. Same code, managed scaling, and free 1:1 WebRTC voice transport.
Inbound or outbound phone calls on Twilio: ConversationRelay if you want Twilio to own speech recognition and synthesis; Twilio Media Streams into LiveKit or Pipecat if you want to choose your own STT and TTS.
Millions of minutes and an in-house media team: Cloudflare Realtime for transport, your own agent workers next to your model provider.
A no-code phone agent instead of a platform: see Retell vs Vapi vs Bland.

How to test a provider in an afternoon

Deploy the provider’s starter agent in the region nearest your users and your model.
Log four timestamps per turn: end of user speech, final transcript, first LLM token, first audio byte out.
Run 100 turns from a real phone and a real mobile network, not only from a laptop on office Wi-Fi, and compare the 95th percentile, not the average.
Price one minute end to end: platform minute + telephony + STT + LLM + TTS. LiveKit’s calculator showed about $0.048 a minute for a sample web stack on October 20, 2026.

Last verified: October 20, 2026.

The short answer

The comparison

Where the 300 milliseconds go

Picks by job

How to test a provider in an afternoon

Sources