A normal call. A lost customer. And a business owner left looking bad.
This happens more often than people think. A customer calls a dealership or repair shop. He speaks fast, the way people actually speak here. He asks: “Do you have a used Camry in stock?” Not a complex request. Not rocket science. Just someone wanting to know if there’s a used Toyota Camry available.
The AI voice agent hears something else. It transcribes “Do you have a used cammery in stock”. The system fails to connect that to Camry, doesn’t map it to inventory, misses the right follow-up question, and goes off the rails. It answers sideways. Something like: “I can help you with used vehicles in general,” or worse, it asks the customer to repeat himself three times. The caller didn’t phone in to test a prototype. He wants a straight answer. So he sighs, hangs up, and calls your competitor.
On the business side, it’s even more frustrating. Because on paper, everything looked good. The vendor promised a “multilingual,” “natural,” “state-of-the-art” voice agent. Nice deck. Nice demo set-up. But in the real world, when someone from Quebec says “Camry” with a local accent, when someone says “I’m checking for a used one” in actual spoken Quebec French, the system falls apart. And then your business looks amateur—not the tech vendor.
Let’s be clear. The problem isn’t that AI is useless. The problem is that Quebec French is a hard mode case for real-time speech recognition. And until very recently, most tools were trained mostly to understand standardized French, often much closer to Paris than to Trois-Rivières.
The good news? On April 27, 2026, ElevenLabs shipped an update to Scribe that genuinely changes the picture for Quebec SMBs. Not a miracle. Not magic. But finally, a concrete technical lever.
Why Quebec French is the difficult case
Spoken Quebec French on the phone is rarely textbook French. It’s alive. Contracted. Fast. Full of elisions, shifting liaison patterns, stronger nasal vowels, integrated English borrowings, trade-specific terms, and brand names pronounced our way. For a local human, no problem. For a speech-to-text engine trained mostly on France French, things can go sideways fast.
Take elisions. A caller doesn’t always say the equivalent of “I am” in full. They contract it. They don’t always say “you know” in a standard form either. They compress, swallow, and blend. If the model expects cleaner forms, it can mis-segment the whole sentence.
Then there’s liaison and chaining. In Quebec, phone speech is often glued together, reduced, compressed. A sentence equivalent to “Do you have any in stock?” may come as one continuous sound block. A phrase like “I want to book an oil change appointment” may mix French syntax with an English-derived verb. If the engine doesn’t know local patterns, it cuts the words wrong or swaps them for phonetically similar nonsense.
Nasal vowels also make things harder. Camry, tires, appointment, morning—on the phone, with a cheap mic or background noise, subtle differences become statistical noise. Quebec’s pronunciation of R, which varies by region, doesn’t help with proper nouns or product names either.
Add commercial anglicisms to that. People say the local equivalent of book a call, do a set-up, check the starter, ask for an upgrade, do a follow-up. In a salon, they’ll mention balayage, blowout, hair botox. In a restaurant: take-out, pick-up, table for six, gluten allergy, extra sauce. If the model doesn’t know the domain, it guesses. And when AI guesses in transcription, you lose calls.
The root issue is training. Models like Whisper have long been strong multilingual baselines, but they were fed massive datasets where France French is overrepresented. The same pattern exists with many newer engines. ElevenLabs Scribe does better on French, and their French speech-to-text page shows that effort, but without contextual adaptation, spoken Quebec business French remained slippery.
Common failures we still see in integrations:
- “Camry” becomes misspellings like cammery or other phonetic variants.
- “oil change” gets turned into unrelated words that sound similar.
- “starter” is warped, dropped, or replaced.
- “book a morning appointment” gets chopped into fragments where only “appointment” survives.
- Local trade terms get treated like names, noise, or filler.
The worst part is this: even when the spoken response sounds great, the intelligence underneath may be running on bad transcription. Then you end up with an agent that speaks like a pro but understands like an exhausted intern. If you want to dig into the “brain” layer behind that, read which LLM to choose, because speech recognition is only one piece of the stack.
What changed on April 27, 2026
On April 27, 2026, ElevenLabs added two features to Scribe Realtime that look modest on paper but are huge for Quebec: keyterms and noVerbatim. You can track that in the official ElevenLabs changelog. And no, this isn’t just a developer footnote. It’s exactly the kind of control that was missing if you wanted to stop getting wrecked by local vocabulary.
Keyterms let you send up to 50 terms, each with a 20-character maximum, directly through WebSocket query parameters. In practice, you tell the engine: “In this call, pay special attention to these words. If what you hear is close, bias toward them.” It’s not a hard override. It’s not magic. It’s contextual biasing.
noVerbatim is a boolean flag that tells the engine not to transcribe every hesitation and spoken crutch word-for-word: filler sounds, false starts, repeated fragments. For a voice agent, that matters. If your pipeline then sends the transcript into a conversational model or intent engine, cleaning those artifacts often improves overall understanding.
Why is this such a big deal for Quebec French? Because instead of hoping the model guesses your world, you can brief it before each call. Your business name. Your products. Your services. Your trade vocabulary. Even frequent customer names when that makes sense. A garage can push Camry, Corolla, starter, oil change, winter tires. A salon can push balayage, stylist names, service names, local city terms. A restaurant can push signature dishes, neighborhood names, room names, combo names.
In other words: stop asking the engine to guess blind. Give it business context before the customer even speaks. That’s where this becomes serious.
Simple technical example:
{
"url": "wss://api.elevenlabs.io/v1/scribe/realtime?language=fr&noVerbatim=true&keyterms=Camry,Corolla,starter,oil%20change,winter%20tires,local%20term",
"headers": {
"xi-api-key": "YOUR_API_KEY"
}
}
Conceptually, that’s it. Real integration still requires handling incoming audio, buffering, segmentation, reconnect logic, and wiring the result to the conversational layer. If you want to compare the real-time conversation side of the stack, OpenAI also laid out its approach with GPT Realtime. But again, without a good transcript up front, even the best LLM can answer the wrong thing.
The key point is this: keyterms finally give you leverage over local accent and niche vocabulary. Not perfect. But concrete. And combined with real business knowledge, it changes production quality—not just demo quality.
What this changes in practice for a garage, salon, or restaurant in Quebec
This starts paying off when we stop talking in abstracts. An AI voice agent for an SMB is not a lab toy. It has to understand your actual business. A garage doesn’t sound like a salon. A salon doesn’t sound like a neighborhood restaurant. With keyterms, we can finally prepare the engine for the real field.
Example brief for a garage:
- oil change
- beater car / local nickname
- starter
- winter tires
- morning appointment
- tune-up
This is not just a cute list. It is a comprehension bias. When the customer says the local equivalent of “I think the starter on my old car is dying”, the engine is more likely to produce a useful transcript instead of a pile of syllables. Then the agent can ask the right follow-up: “Does the vehicle fail to start completely, or only sometimes?”
Mini-case 1: garage in Trois-Rivières
Before the update, a typical call might come through as: “Hi I’d like to book an appointment for my cammery, I think the starter is failing”. The word Camry failed, the booking intent was only partially captured, and starter drifted. Result: the agent didn’t map the request to the right service path and transferred unnecessarily.
With keyterms, we push Camry, starter, appointment, winter tires, oil change. Suddenly the transcript is clean enough for the agent to collect the right details: vehicle year, urgency, preferred time slot, phone number. Nothing flashy. Just useful. That’s what an SMB wants.
Mini-case 2: salon in Sherbrooke
A client calls and says the equivalent of: “Hi, I want to book a balayage with Melanie Saturday morning if there’s room.” Without context, the engine may hesitate on balayage as a salon technique versus a generic word. It may also miss the stylist name or requested time.
With keyterms configured for the salon—balayage, Melanie, women’s cut, root touch-up, Saturday morning—the system understands that this is not vague. It is a booking request for a specific service, with a specific staff member, in a specific time window. If noVerbatim is enabled, filler words also get cleaned up, which helps the booking layer downstream.
The same principle applies to restaurants. If you serve dishes with unusual names, if your neighborhood streets are often misheard, if your callers switch between French and English, keyterms give the engine anchors. After that, you still have to test in the real world—not just in a silent office. We cover that here: test before deployment.
The limits — because no, it’s not magic
Let’s say it plainly: keyterms do not fix everything. This is a major improvement, not a magic wand. If someone calls from a bad line, with the radio on in the car, engine running, kids yelling in the background, you will still get bad audio. No vendor beats physics.
Very strong regional accents are still hard. Calls from parts of Saguenay or Gaspésie, especially with fast delivery and dense local vocabulary, can still trip the engine. Even with good keyterms, if pronunciation drifts too far from what the model has seen, it may still choose the wrong phonetic neighbor.
French↔English code-switching also remains imperfect. A caller who mixes both languages heavily in one sentence is still a challenge. Yes, the system survives better with injected terms. No, it does not become perfect overnight.
Another honest point: there can be a slight latency increase depending on the pipeline, how you push context, and what you do next with the transcript. In a voice agent, a few hundred extra milliseconds can be noticeable if orchestration is sloppy. That’s why integration matters as much as the engine itself.
So yes: it’s better. Clearly better. But you still need design, testing, adjustment, and production monitoring.
Why this changes the math for Quebec SMBs
Before this update, many Quebec SMBs were stuck with a bad choice. Either you used a voice agent that sounded good—nice voice, natural cadence—but understood badly when real Quebec callers phoned in. Or you kept a human receptionist at $28/hour or more, with breaks, dead time, missed evening calls, and staffing headaches.
Now the cost/quality ratio starts to shift for real. Not because AI suddenly became perfect. Because we finally have a practical way to reduce the gap between how Quebec customers actually speak and the kind of transcript the system can use. When you combine keyterms, noVerbatim, a solid knowledge base, clean prompts, and proper CRM or calendar integrations, you get something much more serious than a trade-show gimmick.
This is exactly where TECHMA comes in. We do not throw you a dashboard and tell you to figure it out. We configure everything: keyterms, prompts, call flows, integrations, load tests, edge cases, human transfers. Because an AI voice agent in Quebec French is not something you deploy self-serve if you want it to actually hold up.
That’s the real shift: for a Quebec SMB, the voice agent finally becomes a realistic option for handling repetitive calls without sounding incompetent on the phone.
TECHMA configures everything. You answer the phone less.
We do not sell software. We deploy a complete solution configured for your industry, your vocabulary, your customers, and your real operational constraints. Garage, salon, clinic, restaurant, dealership—we adapt the agent to your world, not the other way around.
If you want to see what an AI voice agent for Quebec French looks like when it’s built properly, we can show you in a 30-minute demo. Not a fake demo with perfect calls. A real discussion about your case, your words, your integrations, and what the April 2026 ElevenLabs update changes for you in practice.
Book a free 30-minute demo with TECHMA. We handle the set-up, integration, testing, and fine-tuning. You recover calls you were losing.