On May 4, 2026, OpenAI published a technical post on how they rebuilt their WebRTC infrastructure to serve 900 million weekly users in under 500ms round-trip. Most Quebec SMBs missed it — and yet it explains exactly why your voice AI agent either has a shot at sounding human on the phone, or has none at all.
The catch: nearly every voice agent provider lies about their numbers. They advertise "75ms latency" on the homepage, but once the system is running in your business, customers hang up after two seconds of silence because they think the line dropped. Let's break it down.
The 300ms threshold: why it's neurological, not marketing
In a normal human conversation, the gap between the end of one person's sentence and the start of the other's reply sits around 200 to 300 milliseconds. That's not an averaged-out marketing number — it's brain wiring. Past 300ms, your conversation partner unconsciously senses a delay. Past 500ms, they consciously notice. Past one second, the brain logs that something is off.
For your Quebec customers calling your SMB on a Monday morning, that threshold is non-negotiable. If your voice agent takes two seconds to say "Hello, you've reached..." the caller has already started saying "hello? hello?" over top. Now the agent has to handle an interruption while it was just starting to speak. Result? Garbled audio that sounds robotic, and a customer who hangs up.
The concrete numbers, for context:
- Under 800ms round-trip: conversation flows naturally, the customer doesn't notice anything.
- 800ms to 1.2 seconds: acceptable for a standard business call, but the customer feels it's "robotic."
- 1.5 to 2 seconds: the customer wonders if the line is still active. Abandonment rates climb 40%+.
- Beyond 2 seconds: the conversation breaks. The customer hangs up or gets frustrated.
The classic lie: "our agent answers in 75ms"
Browse most voice agent provider websites. You'll see impressive numbers: 75ms, 100ms, sometimes "sub-100ms." The catch is those numbers measure only one stage of the full pipeline. It's like saying "my car does 0-60 in 5 seconds" while measuring only the moment the clutch engages.
The real pipeline of an AI voice agent phone call has at least five stages:
- Audio capture + transmission to server: 50-150ms depending on Quebec network quality (cell service in Lac-Saint-Jean isn't downtown Montreal).
- Speech recognition (STT): 100-400ms to turn voice into text.
- LLM inference: 200-800ms for the model to generate a response — this is the most variable stage depending on which model you're using.
- Voice synthesis (TTS): 100-300ms to turn text back into voice.
- Return to the customer's phone: 50-200ms.
End-to-end total in a typical poorly-configured Quebec SMB? Between 1.5 and 3 seconds. That's exactly why your customers say "voice agents sound robotic." It's not the synthetic voice — synthetic voice has gotten excellent. It's the silence before the voice.
SignalWire published a fairly direct piece on this recently, worth reading if you're shopping for a solution: their analysis confirms that real-world production latency averages 2 to 3 seconds, not 75ms. And platforms like LiveKit now display latency in real time during calls — you can see it for yourself.
What OpenAI's May 4 announcement actually changes
On May 4, 2026, OpenAI's infrastructure team published a technical post explaining how they rebuilt their WebRTC stack to hit 300 to 500ms round-trip for the majority of sessions, at global scale. Three changes that matter for you in Quebec:
1. Edge routing instead of datacenter routing
Before, your customer called from Sherbrooke, the signal traveled to a US datacenter (typically Virginia or Oregon), then back. That's an easy 80-150ms of pure network latency for the round trip. Now, OpenAI routes the session to the closest media router to the customer — for Quebec, likely a Toronto or New York node. Net difference: 30-50ms saved per turn of speech.
2. "Relay + transceiver" architecture
Instead of having a full WebRTC session locked to a single port for the entire call (the old model), OpenAI now terminates the connection at the edge and converts audio into simpler internal protocols. Concretely: less jitter, fewer dropped packets, and a connection that establishes faster when your customer starts speaking.
3. Optimized one-to-one sessions
OpenAI confirms that the majority of sessions are 1-to-1 — one human, one agent. Their entire architecture is now built around that use case, which is exactly yours when a customer calls your clinic, your law firm, or your restaurant. The optimizations they shipped directly benefit SMBs running on the Realtime API.
You can read the technical detail directly on OpenAI's blog — it's dense, but the conclusion is clear: the bar just moved up, and solutions still sitting at 1.5-2 seconds in 2026 are going to look generation-old by August.
How to verify your own agent's real latency (in 10 minutes)
Here's what we do at Agent IA Vocal to measure the real latency of an agent we deploy at a customer site:
Test 1: timed recording. Call your own number, record the call (with consent, obviously), and play it back in an audio editor like Audacity. Measure the time between the end of your last syllable and the start of the agent's first syllable. Do this over 10 turns. The average is your real round-trip latency. If it exceeds 1.2 seconds, you have a problem.
Test 2: the "uh-huh" test. During a test call, speak short sentences with clear endings ("okay.") and note whether the agent talks over you. A well-configured agent waits until you've finished, but jumps in within the 300-600ms window. Too early = it cuts you off. Too late = there's a silence gap. The setting that controls this is usually called "endpointing" or "turn detection," and we've written a full guide on what goes wrong when it's misconfigured.
Test 3: the multi-network test. Test from three different locations: your office wifi, your LTE cell, and ideally a cellular line in a more rural area (an employee working from home in Saint-Hyacinthe, say). If latency varies by more than 500ms between the three, your solution doesn't have proper edge routing — and customers calling from area code 819 or 418 are getting a noticeably worse experience than the 514 ones.
The real questions to ask your provider this week
If you're currently evaluating a voice agent, or you have one that "sort of works," send these four questions by email to your provider:
- "What is your end-to-end round-trip latency, measured from the moment the customer finishes speaking to the moment the agent's first syllable hits the phone — not an isolated pipeline stage?"
- "What WebRTC or SIP infrastructure handles the call, and where is the closest physical media termination point to Quebec?"
- "Have you migrated to OpenAI's new WebRTC architecture announced May 4, 2026, or are you still on the old one? If not, what's the timeline?"
- "Can you show me a recorded test call with timestamps for each stage (STT, LLM, TTS) so I can see where the time goes?"
If the provider deflects on questions 1 or 4, you have your answer. A serious provider has those numbers and can deliver them within 24 hours.
What this means for your LLM choice
The LLM's latency (stage 3 of the pipeline) is the most variable stage, and the one where you have the most leverage. Full GPT-4o averages 600-800ms. GPT-Realtime-Mini, launched in April 2026, can drop to 200-300ms for simple replies. Gemini 3.1 Pro Preview sits around 400-500ms. A well-hosted open-source Qwen 35B: 300-400ms.
For your SMB, the right choice depends on call complexity. A vet clinic taking simple appointments? Mini is plenty, and you save 70% on your API bill on top. A law firm handling nuanced questions? Full GPT-Realtime, because precision matters more than the microsecond. We did a detailed comparison of the three main voice agent LLMs for Quebec that goes into the use-case detail.
The trap: optimizing latency at the expense of quality
Watch out, though. We regularly see SMBs fall into the opposite trap: they switch to the fastest model (often a Whisper-Tiny for STT and a mini-LLM), drop down to 600ms round-trip, and three weeks later realize the agent badly misunderstands Quebec French. The customer says "mon char est au garage" and the agent hears "mon chat est au garage" (car vs cat). You gain 400ms of latency, you lose 15% of comprehension accuracy. Bad trade.
The right reflex is to optimize in this order:
- First, STT quality in Quebec French — without it, nothing else matters.
- Then LLM latency, picking the smallest model that holds accuracy on your real use cases.
- Finally, turn detection and pre-tool speech configuration — which can save 500ms of perceived delay without touching the pipeline. We wrote a precise guide on this after the April 27 ElevenLabs update: how to fix the dead air that's killing your calls.
And what about OpenAI's direct SIP?
Quick reminder for those who missed the previous episode: on April 30, 2026, OpenAI also launched a direct SIP connection to GPT-Realtime — meaning a phone call can now enter the model directly without going through a traditional WebRTC layer. Combined with the May 4 infrastructure rebuild, this means that for a Quebec SMB taking calls through Twilio or a Quebec SIP trunk, the path can now be shorter than before. We detailed what direct SIP changes concretely in another article.
Combining both announcements (direct SIP + rebuilt WebRTC): we're talking about a realistic gain of 200 to 400ms round-trip for a typical configuration. That's not minor. That's the difference between an agent that sounds robotic and an agent that goes unnoticed.
The 2026 gold standard: where your agent should be in 90 days
Here's the grid we use to qualify a Quebec SMB voice AI agent in May 2026:
- Excellent: end-to-end round-trip latency under 800ms, measured over 50 real calls from three different Quebec regions. The customer doesn't know they're talking to AI.
- Acceptable: 800ms to 1.2s, with well-configured pre-tool speech that masks the pauses during API calls (booking, calendar checks).
- Needs rework: above 1.5s. Your customers are already telling you, even if they don't articulate it in technical terms. They say "it sounds weird" or "I dunno, I just don't like it."
- Catastrophic: above 2s. You're losing money on every call.
If your agent is in the "needs rework" or "catastrophic" zone, the May 4 announcement is your way out. An architecture overhaul that looked expensive six months ago is now standard at serious providers. It's the right time to ask for a rebuild.
The verdict for your Quebec SMB in May 2026
Latency isn't a technical detail. It's the variable that decides whether your voice agent sounds human or robotic. And with what happened between April 30 (OpenAI direct SIP) and May 4 (WebRTC rebuild), the bar just moved up. Solutions that were acceptable in January 2026 are a generation behind in May.
If you run a Quebec SMB — clinic, law firm, restaurant, retail, real estate agency — and your voice agent has been live more than six months without an infrastructure review, it's the moment to measure its real latency and push your provider to upgrade. Otherwise, the competitors doing this work this week will build a lead that's hard to close by fall. At Agent IA Vocal, the Techma team handles the entire setup, integrations, and migration to the new infrastructure for you — there's nothing for you to configure yourself.