Phone Call Audio Streaming to an LLM Under 300ms: Providers
The short answer
For a phone agent, pick the provider that owns the call, then keep speech recognition, the model and speech synthesis as close to that network as you can. As of October 28, 2026:
- Telnyx Voice AI — the fewest hops: carrier, speech-to-text, text-to-speech, orchestration and LLM inference on Telnyx’s own network for $0.05 a minute plus tokens and carrier cost. Best for new phone agents where latency matters most.
- Twilio Media Streams / ConversationRelay — the largest telephony estate and documentation; stream raw call audio to your own stack ($0.0044 a minute) or let Twilio handle speech and send your LLM text ($0.07 a minute).
- Plivo audio streaming — the cheapest raw minutes: streaming and noise cancellation included, inbound local calls $0.0055 a minute.
- LiveKit Cloud telephony — best when the same agent also serves web and mobile users: SIP and phone numbers into LiveKit rooms, $0.01 a minute telephony plus $0.01 per agent-session minute.
This page is the phone-call spoke of our cloud providers for real-time audio streaming to LLM apps comparison, which covers web and mobile transport.
The comparison
Prices are US list rates from each vendor’s pricing page on October 28, 2026.
| Provider | How audio reaches your LLM | Call price (US) | AI layer price | Best for |
|---|---|---|---|---|
| Telnyx Voice AI | Hosted agent runtime; or Media Streaming over WebSockets ($0.0035/min) | SIP inbound from $0.0032/min, outbound from $0.005/min; numbers $1/month | $0.05/min voice engine (STT + TTS + orchestration); LLM per token on Telnyx GPUs | Lowest-latency all-in phone agents |
| Twilio | Media Streams (bidirectional WebSocket) or ConversationRelay (text to your LLM) | Receive local $0.0085/min + $1.15/month number; make $0.0140/min | Media Streams $0.0044/min; ConversationRelay $0.07/min | Existing Twilio estates, compliance paperwork |
| Plivo | Audio streaming over WebSocket | Receive local $0.0055/min; make $0.0115/min; numbers $0.50/month | Streaming and noise cancellation included | Cost-sensitive high volume, bring your own STT/LLM/TTS |
| LiveKit Cloud | SIP / phone numbers into a room with a hosted agent | Telephony $0.01/min; 1 free number on Build | Agent session $0.01/min; LiveKit Inference or your own keys | One agent for phone, web and mobile |
Where the 300 milliseconds go on a phone call
A phone call adds two costs a web session does not have: the public telephone network’s own delay, and narrowband audio. Twilio’s Media Streams documentation states the stream is always audio/x-mulaw at 8,000 Hz, mono, base64-encoded, and audio you play back must be in the same format. Budget from the moment the caller stops speaking to the first audio byte they hear:
- End-of-turn detection — the agent decides the caller has finished. Too eager and it interrupts; too slow and the call feels laggy.
- Final transcript from streaming STT trained or tuned for 8 kHz phone audio.
- LLM time to first token — the biggest single item. Use a small, fast model and prompt caching; see our lowest-latency LLM inference frameworks.
- TTS time to first audio, then transcoding back to μ-law.
- Network legs between the carrier’s media server, your agent and each model API. Every extra provider is another round trip; this is Telnyx’s argument for one network.
A speech-to-speech model removes steps 2 and 4 as separate services. Per-minute rates for GPT-Live-1, Gemini 3.8 Live and Grok Voice 3.0 are in our current API prices.
Provider notes
Telnyx bundles orchestration, hosted STT (Deepgram or Telnyx STT) and a choice of TTS voices into one $0.05 rate and bills LLM tokens separately at GPU rates; its own estimate for a production agent is about $0.06 a minute. Two caveats from its pricing page: the voice engine is rounded to 60-second increments per call, which hurts short calls, and the pay-as-you-go plan allows 500 concurrent calls. The sub-200 ms response figure is Telnyx’s claim.
Twilio gives you two paths. Media Streams forks or replaces call audio over a WebSocket to your server, so you choose every model. ConversationRelay has Twilio run speech recognition and synthesis and exchange text with your LLM — simpler, but you give up choice of STT and TTS.
Plivo includes audio streaming and noise cancellation in the call price and has the lowest US minute rates of the four, which matters at hundreds of thousands of minutes a month.
LiveKit treats a phone call as another participant in a WebRTC room, so the same agent code, turn detection and observability serve phone and app users. Its calculator showed about $0.048 a minute for a sample phone stack with telephony, STT, LLM, TTS and observability.
How to test in a day
- Place 100 calls from real mobile phones, not a softphone on office Wi-Fi.
- Log end of caller speech, final transcript, first LLM token and first audio byte for each turn; compare the 95th percentile.
- Run the agent and models in the same region as the carrier’s media edge.
- Price one minute end to end: call + streaming + STT + LLM + TTS. For a no-code phone agent instead, see Retell vs Vapi vs Bland.
Last verified: October 28, 2026.