Update: May 7, 2026, 10:17 AM — less than four hours ago. OpenAI just announced three new audio models in its API. One of them, GPT-Realtime-Translate, translates voice in real-time across 70+ input languages and 13 output languages. And yes, French and English are both on the list.
We spent the morning testing it head-to-head against the solution you're probably already running: ElevenLabs in multilingual mode. Here's what we found for the scenario that literally defines customer service in Quebec — the famous Bonjour-Hi.
The problem we're actually trying to solve
You run a shop on Saint-Denis Street. The phone rings. You answer "Bonjour-Hi!". The other person continues in English. You switch. Three minutes later, their sister-in-law grabs the phone — she only speaks French. You switch again. That's the lived reality of 38% of Montreal SMBs according to Statistics Canada.
Now replace "you" with your AI voice agent. The challenge stays the same: detect the language, switch cleanly, and most importantly don't sound like a robot when the customer mixes both within the same sentence ("Can you check si vous avez ça en stock?"). That's code-switching, and it's an Olympic sport in Quebec.
Until this morning, you basically had two serious approaches: a single ElevenLabs agent in multilingual mode, or two separate agents with a language router in front. With GPT-Realtime-Translate, OpenAI opens a third door. Question is: which one wins for a Quebec SMB?
Approach A: ElevenLabs Multilingual (the current baseline)
ElevenLabs has handled French and English in a single agent since Conversational AI 2.0. You configure one prompt, you flip on multilingual, you pick a voice that sounds good in both languages (Charlotte or Adam, typically), and away you go.
What works well: the voice stays the same when the customer switches. It's coherent, it builds trust, and nobody wonders why "the lady on the phone" suddenly changed timbre. Latency runs around 600-900 ms in Canadian French based on our recent tests — not perfect, but usable. We've covered the latency topic in detail here.
What works less well: mid-sentence code-switching. If the customer says "Je voudrais book un rendez-vous next Tuesday," ElevenLabs picks a dominant language and treats the rest as noise. Result: it asks you to repeat, or it interprets "next Tuesday" phonetically in French and goes off the rails. Manageable, but real friction.
Approach B: GPT-Realtime-Translate (the newborn)
The model OpenAI shipped this morning is different in its DNA. It's not a conversational agent per se — it's a streaming voice translator. You speak French, it comes out in English. With announced latency under 800 ms and pricing of US$0.034 per minute — roughly 4.7¢ CAD per translated minute of conversation.
The scenario where it shines: your SMB has a primary voice agent operating in one language (let's say French), but regularly receives calls in English. Instead of building a second full agent, you slot GPT-Realtime-Translate as a layer between the English-speaking caller and your French agent. The customer speaks English, the agent hears French, the agent answers in French, the customer hears English. Nobody loses time.
According to OpenAI's official announcement, Deutsche Telekom is already piloting this exact pattern for international customer support. That's a serious deployment signal.
What works less well: the output voice isn't your agent's — it's OpenAI's neutral translator voice. You lose your brand's sonic signature. And for now, you can't fine-tune the translator's tone or personality, unlike an ElevenLabs agent you can make significantly more natural with Expressive Mode.
The real test: three Quebec scenarios
Scenario 1 — The clearly anglophone call. An American tourist calls your inn in Tadoussac to book. No mixing, just clear English. Verdict: ElevenLabs wins by a notch. The voice stays the same throughout, it's warmer, and latency is better because there's no extra translation step. GPT-Realtime-Translate adds 200-300 ms of overhead minimum.
Scenario 2 — The clearly francophone call, but the customer briefly switches. A lady from Sherbrooke calls you, speaks French, but slips in "like c'est-tu open le dimanche?". Verdict: tie. ElevenLabs handles it well because context stays mostly French. GPT-Realtime-Translate gets confused because it tries to translate an isolated English word that didn't need translating.
Scenario 3 — Aggressive code-switching (the real Montreal). A customer calls and zigzags through the whole conversation: "Hi, je voulais savoir if you guys do delivery sur le Plateau? Like what's the minimum order?" Verdict: GPT-Realtime-Translate wins, but not by KO. It maintains both language contexts simultaneously better than ElevenLabs, which ends up asking to repeat at least once in every three calls in our testing.
What it actually costs, in Quebec numbers
A typical Quebec SMB taking 200 calls per month (avg 4 minutes/call = 800 minutes):
- ElevenLabs Multilingual alone: roughly $96 CAD/month in API consumption (Conversational AI 2.0 tier).
- ElevenLabs + GPT-Realtime-Translate as a layer: $96 + $38 for the ~800 translated minutes = $134 CAD/month.
- Two separate ElevenLabs agents (FR and EN) with router: roughly $152 CAD/month plus operational complexity.
For an extra $38 per month, the layered option is interesting if your customer base is genuinely 60/40 or more imbalanced. But for an 80/20 French-English SMB, ElevenLabs alone holds up and costs less.
And what about Law 25?
Often-forgotten detail: adding a translation layer means the customer's voice transits through one extra service (OpenAI in this case). If you're handling personal information, that's another entry in your processor registry. Before plugging GPT-Realtime-Translate into an agent that collects sensitive data (clinic, financial, legal), run through our 12-question Law 25 compliance test. This isn't optional.
OpenAI announces European data residency but not Canadian-specific yet. For retail or restaurants, it's fine. For a law firm or a clinic, ask the question twice.
Our practical verdict for your SMB
If your customer base is mostly francophone with a few anglophone calls per week, stay on ElevenLabs Multilingual. Voice consistency and lower cost easily outweigh the rare code-switching glitches. We already recommended ElevenLabs as the Quebec baseline in our TTS comparison earlier this week, and this morning's news doesn't change that starting point.
If you operate in downtown Montreal with a genuinely 50/50 customer base and aggressive code-switching is the norm in your calls, GPT-Realtime-Translate as a layer is worth testing. Plan two weeks of parallel testing with completion-rate measurement before flipping production.
If you operate in rural Quebec (Lac-Saint-Jean, Gaspésie, Bas-Saint-Laurent), forget translation and stay French. The added complexity doesn't justify itself for 2-3 anglophone calls per month.
One last thing: GPT-Realtime-2 also dropped this morning and it's a different model, more oriented toward a full conversational agent. Don't mix the two. Translate translates, Realtime-2 drives the conversation. According to 9to5Mac's coverage this morning, OpenAI positions them as complementary rather than competing.
At Agent IA Vocal, we've already started integrating GPT-Realtime-Translate into deployments for clients who need it — installation, configuration, and Law 25 compliance are handled end-to-end by our TECHMA team. If you'd like to see how it would perform on your real calls, we're taking pilot requests for May and June 2026.
Questions people asked us this morning
Does GPT-Realtime-Translate replace my ElevenLabs agent? No, and it's not designed to. It's a translator, not an agent. It doesn't follow a script, doesn't query your knowledge base, doesn't take reservations. It transforms voice from one language to another. If you want to replace ElevenLabs with full OpenAI, look at GPT-Realtime-2 instead, which we covered in a separate article this morning.
How good is the English output voice? Honest but generic. No strong accent, no particular warmth, no Quebec intonation. Sufficient for functional but it won't sell your brand. For brands where the voice IS the product (premium call centers, luxury real estate), ElevenLabs remains superior.
How long to plug this into an existing agent? For a clean deployment with testing, plan 8 to 12 hours of technical work. Less than you might think because the API is very well documented, but the longest part is testing with real calls and real Quebec accents. Don't skip this step.
What about joual? What about chiac? Preliminary tests: GPT-Realtime-Translate handles standard Quebec French without issue. Very regional expressions ("j'aurais besoin de checker ça à matin") translate word-for-word giving correct English but without the naturalness. For Acadian chiac, we don't have enough data yet. To be monitored.
Can I disable translation for certain numbers? Yes, the architecture allows conditional routing. If you have a dedicated francophone line and another bilingual one, you can activate Translate only on the second. That's exactly what we configure for the bilingual dental clinics in Laval we work with.
The bottom line
OpenAI just placed a new piece on the chessboard this morning. It doesn't change the whole game, but it opens an option that didn't exist yesterday. For the typical Quebec SMB, ElevenLabs Multilingual remains the healthy baseline. For genuinely bilingual operations in downtown Montreal, evaluating the Translate layer becomes serious. And for regions, stay French — your customer base will thank you.
The good news: regardless of the scenario, the technical complexity no longer lands on your shoulders. That's exactly what we exist for.