Both panels talk to the same OpenAI Realtime model. The left is a typical tutorial integration over WebSocket. The right is the full Telequick stack: QUIC transport, warm OpenAI session pool, and a tuned bargein path. Try interrupting the agent on each — the difference you hear is the gap between a stack you build over a weekend and one we've spent months tuning.
Try one or both calls below, then copy the recorded numbers and paste them into any LLM for a second opinion on which transport performed better.
The difference you hear when you interrupt is the gap between a stack you'd assemble from an OpenAI Realtime tutorial and what we ship in production — including a warm OpenAI session pool that keeps connection setup off the critical path, and a bargein path tuned end-to-end so the agent stops the instant you start speaking.
On a clean network, transport alone (TCP vs QUIC) buys you a tighter histogram tail under packet loss. To see the transport story by itself, throttle the network in DevTools and rerun both panels — the post-call summary's inter-arrival histogram is where it shows up.
WebSocket to OpenAI Realtime with default turn-taking. What you'd ship after a tutorial — agent tails off for a beat before going silent.
Same model, routed over QUIC with a warm OpenAI session pool and a tuned bargein path. The agent stops the moment you start speaking.
Talk over the agent while it's mid-sentence, mid-paragraph, or mid-30-second monologue. Voice cuts in well under a second — the moment you start talking, not the moment OpenAI's server-side VAD catches up.
Interrupt, agent stops, agent replies, interrupt again. No truncation queue, no stuck audio gates, no 'now the next reply is silenced' bug — every barge-in re-enters cleanly.
A warm OpenAI session pool keeps connections pre-dialed and pre-configured, so the agent's first words land without the awkward setup gap a fresh dial would add.
QUIC's per-stream framing means a single dropped packet doesn't head-of-line-block all your audio. Throttle your network to 3% loss in DevTools — the inter-arrival p95 widens far less than on TCP.
getUserMedia echo cancellation + a backend cooldown gate stop the agent's own audio bleed from tripping VAD as if the user spoke. No 'constant interruption' loop on speakers.
Chrome / Edge get QUIC. Safari, Firefox, locked-down corp browsers transparently fall back to WSS with the same agent and the same fast-cancel bargein path.
Copies a markdown summary of both calls and a short prompt. Paste into ChatGPT, Claude, Gemini, or whichever model you prefer.
Seastar shard-per-core lets a single commodity bare-metal host (8 vCPU, 16 GB RAM, ~$200/month) carry ~2,500 concurrent voice-AI calls — roughly 30M minutes of monthly traffic at typical 30% daily-curve utilization. Numbers below project what each managed vendor would cost to carry that same envelope.
| Vendor | $/min | 100k min/mo | 1M min/mo | 10M min/mo | 1 box (30M min/mo) | vs Telequick |
|---|---|---|---|---|---|---|
Telequick (Seastar + QUIC) Bare-metal C++23 shard-per-core, QUIC media. Infra cost only. | $0.001 | $100 | $1.0k | $10k | $200 | baseline |
LiveKit Cloud SFU per-participant-minute + bandwidth, agent SDK overhead. | $0.005 | $500 | $5.0k | $50k | $150k | 5× |
Twilio Voice + Media Streams PSTN min + Media Streams egress. PCMU only. | $0.014 | $1.4k | $14k | $140k | $420k | 14× |
Vapi Platform fee on top of underlying carrier + model spend. | $0.050 | $5.0k | $50k | $500k | $1.5M | 50× |
Sources: livekit.io/pricing, twilio.com/voice/pricing/us, vapi.ai/pricing. Telequick capacity is measured for audio-transport only (Opus 24 kHz, QUIC, no co-located inference). Real workloads with on-box STT/TTS/LLM will reduce the concurrent-session figure proportionally to inference cost. Talk to us for a workload-specific sizing.
Short reads on where Telequick differs from LiveKit, Vapi, Twilio Media Streams, and the WebRTC platform pack. Built for engineers evaluating which transport layer to bet on for voice agents.
LiveKit is a general-purpose WebRTC SFU adapted for voice agents. Telequick is purpose-built for sub-100ms voice AI on QUIC.
Vapi is an opinionated voice-agent platform that abstracts the stack. Telequick is the transport layer underneath, with no markup.
Twilio Programmable Voice + Media Streams was the default for voice AI before WebRTC. Telequick replaces the Media Streams hop.
Other WebRTC platforms share LiveKit's general-purpose tradeoffs: built for video conferencing first, voice agents second.