Friday 6:45 PM, rue Saint-Denis. Your hostess is juggling three iPads, two empty water carafes and a couple at table 4 still hesitating between the tartare and the octopus. The phone rings. Nobody picks up. On the other end, someone wanting to book a table for four in Mandarin for Saturday night. They hang up after four rings. They open OpenTable. They book your neighbor instead.
That scene isn't an exception in Montreal anymore. According to the OQLF's 2024 report on the linguistic situation in Quebec, 50% of customers in the Montreal metropolitan area report having been served at least once in a language other than French in 2018 — up from 25% in 2010. On the Island of Montreal, allophones now make up 32.5% of the population. For you as a restaurateur, that means roughly one call in three — sometimes one in two in certain neighborhoods — does not arrive in your host's strongest language. And most of those calls hit at the worst moment: right before service.
The quick math that should hurt
Quick exercise. Average cover in a Montreal neighborhood bistro: around CA$50. Party of four: CA$200. If you lose even one allophone reservation per night because nobody picks up or because the host can't understand the request, that's CA$1,400 per week. Annualize that and you're looking at over CA$70,000 evaporating into thin air. And that's the conservative estimate — for many restaurants in the Mile-End, Chinatown or Côte-des-Neiges, the real number is materially higher.
Good news: between May 4 and May 12, 2026, two major changes landed in the voice agent ecosystem. ElevenLabs added gemini-3.1-pro-preview to its list of available brains in its May 4, 2026 changelog, and OpenAI shipped a family of three new realtime voice models on May 7: gpt-realtime-2 (GPT-5-class reasoning, 128,000-token context), gpt-realtime-translate (live translation across 70+ source and 13 target languages), and gpt-realtime-whisper (streaming transcription).
The direct consequence for a Montreal restaurant: you now have three different technical stacks to build an AI voice agent that takes reservations in your allophone customer's language. And those three stacks have very different cost, latency and Quebec French voice quality profiles. Here's what to know before signing anything.
Before the comparison: the confusion to kill right now
Many articles published this week conflate two very different things. gpt-realtime-translate is NOT for building agents. OpenAI says so explicitly in the documentation: this model is designed for live translation between two humans, not for a bot to make decisions, call tools, or save a reservation. If someone shows you a demo where "gpt-realtime-translate takes the reservation," run. You'll have a translator that repeats the customer's words in French, not an agent that checks the OpenTable calendar and confirms a table for four at 7:30 PM.
The agent that makes the decision must be gpt-realtime-2 — or an ElevenLabs agent with an LLM brain (Claude, GPT, or now Gemini 3.1). gpt-realtime-translate, however, can be useful in some hybrid architectures to translate menu content into the caller's language while the agent maintains the conversation. It's useful, but secondary.
With that cleared up, here are the three stacks that actually work for your restaurant.
Stack A — ElevenLabs Agent with multilingual TTS (the 2026 standard)
This is the stack most Montreal restaurants TECHMA has deployed this year are running. The AI voice agent runs on the ElevenLabs Agents platform (same one that received the text-behavior-overrides update on May 25, which we analyzed last week). The brain can be Claude Sonnet 4.5, GPT-4.1, or — since May 4, 2026 — Gemini 3.1 Pro Preview. The outbound voice uses ElevenLabs' v3 multilingual models, which speak 29 languages with the same vocal fingerprint.
How it works in practice. A customer calls in Mandarin. The agent detects the language within the first second, switches the TTS to the Mandarin voice of the same persona (as we explained for Montreal Med Spas in May), and carries the conversation. The brain LLM (often Claude for this case) reads the restaurant context, checks availability via a Resy or Libro Reserve API, suggests a slot, takes the SMS confirmation, and sends a summary into Zoho for post-visit marketing.
Typical latency: 800 to 1,100 ms from the customer's last word to the start of the voice reply. Enough to not break the flow.
Cost: between CA$0.12 and CA$0.18 per minute depending on the chosen LLM. An average reservation conversation runs around 90 seconds, so ~CA$0.20 per call answered. For 600 calls per month, you're looking at CA$120 all-in.
The weak spot. Quebec French quality is decent but not perfect — the default accent leans European French. For a bistro serving a largely demanding francophone clientele (Outremont, Plateau), it's a point to tune with a custom Quebec voice clone. That's part of the configuration work the TECHMA team handles for each client during deployment.
Stack B — OpenAI gpt-realtime-2 directly (the new-generation heavyweight)
Released May 7, 2026, this stack short-circuits ElevenLabs orchestration. You talk directly to OpenAI's Realtime API, which handles everything: transcription, reasoning, voice generation. gpt-realtime-2 has GPT-5-class reasoning, a 128,000-token context window (up from 32,000 in the December 2024 version), and a 30.5% score on MultiChallenge audio — vs 20.6% for the model it replaces. In plain terms: it understands complex requests better, handles customer interruptions better, and can call multiple tools in parallel.
How it works. A customer calls, speaks in Spanish. gpt-realtime-2 detects the language, responds in Spanish with a native voice (the model ships with 8 built-in voices), queries the reservation tool, confirms the table, and hangs up. No middleman. No routing between providers.
Latency: 400 to 700 ms — roughly 30% faster than Stack A because there's no hop between transcriber, LLM and voice synthesizer.
Cost: US$32 per million audio input tokens, US$64 per million output tokens. For a typical 90-second conversation, around CA$0.28 per call answered. More expensive than Stack A, but the quality jump is tangible on complex requests ("I want to book for 8 but my mother-in-law is in a wheelchair and my nephew has a shellfish allergy").
The weak spot. Less control over voice personality. OpenAI's 8 voices are well-engineered but not customizable to the same degree as cloned ElevenLabs voices. For a restaurant that bets on a strong sonic identity (your signature maître d'), Stack A remains better.
Stack C — The hybrid: ElevenLabs Agent + GPT-Realtime-Translate for the menu
This is the rarest stack, but it best fits the gastronomic restaurant with Asian or European tourists. The agent's core stays ElevenLabs (Stack A), but when the customer asks about a dish — ingredients, allergens, cooking method — a parallel call to gpt-realtime-translate live-translates the menu description (written in French by the chef) into the customer's language.
The win: the menu is always current because the chef writes it in French in a Google Doc, and gpt-realtime-translate renders it in Mandarin, Italian, Arabic or Hindi without anyone having to maintain six translated versions. When the chef changes the sauce of the day, it reflects immediately in every call, regardless of language.
The downside: it's a complex architecture to orchestrate. Webhooks, state management across two providers, monitoring two invoices. Not a do-it-yourself deployment. It's exactly the kind of integration the TECHMA team builds custom when a gastronomic client needs it — not a turnkey product you flip on with two clicks.
4-criteria decision matrix
Here's the table our account managers use with Montreal restaurateurs who are weighing options.
Which choice for which Montreal restaurant type?
Neighborhood bistro (Plateau, Rosemont, Villeray): Stack A. You have a majority francophone clientele but 20-30% allophones, especially at night. The Quebec voice clone makes a real difference on the francophone experience, and v3 multilingual covers the anglophones and Latin Americans in the neighborhood at no extra cost.
Fine cuisine with tourists (Old Montreal, Downtown): Stack C. When 40% of your clientele comes from outside Quebec and your menu changes three times a week, paying CA$0.07 more per call for an auto-translated menu in 13 languages saves time and prevents allergen mistakes that can cost a customer's life.
Fast-casual chain (Olive Group, MTY, Subway franchises): Stack B. The volume justifies the extra spend because of lower latency (customers hang up less often) and the reasoning robustness when routing to the right location among 14.
Food truck and sandwich shop: none of the three. Your call volume doesn't justify a voice agent. Put your energy on Instagram and Foodora/Uber Eats.
Law 25: the detail nobody covers
All the stacks above record the conversation (transcript and sometimes audio) for quality and compliance. In Quebec, that falls directly under Law 25 on the protection of personal information. Concretely: your agent must announce the recording in the first seconds (in French, and in the customer's detected language if different), you must have a named privacy officer, and transcripts cannot be transferred out of Canada without a documented risk assessment.
None of the three providers (ElevenLabs, OpenAI directly) handles that for you. It's the TECHMA team that configures the bilingual disclosure banner, sets the access controls in Zoho or your CRM, and writes the Privacy Impact Assessment required by the CAI. Without it, your deployment isn't compliant — and the fine can reach CA$10 million or 2% of worldwide revenue.
What to take away
Three May 2026 changes have reshaped the landscape: Gemini 3.1 available in ElevenLabs, gpt-realtime-2 twice as performant as its December 2024 predecessor, and gpt-realtime-translate democratizing live translation (but not for agents — for humans). For a Montreal restaurateur seeing 30% of calls arrive in a language other than French, doing nothing is expensive: every Friday without an agent is roughly CA$800 going to the competitor next door.
Stack A for most neighborhood restaurants. Stack B for high-volume chains. Stack C for tourist-driven gastronomy. None of them deploys in two clicks — which is precisely why the TECHMA team handles integration end-to-end, from picking the Quebec voice to Law 25 compliance to the webhooks into OpenTable, Libro or your Toast/Lightspeed POS.
Want to test which of the three stacks fits your restaurant profile? Book a demo with the team — we take 20 minutes to analyze your current call volume, your real linguistic mix, and the cost of a missed call in your specific operation. It's free, and you leave with a concrete number, not a marketing promise.