For years, there was a quiet tell that gave away an AI voice agent the moment it switched languages. You'd call a business, the English greeting would sound crisp and confident, and then — the second a French-speaking caller said bonjour — the voice would flatten into that stiff, over-enunciated cadence we all recognize. The machine could speak French. It just couldn't sound like it lived there.
That tell is disappearing, and faster than most business owners realize.
On September 28, 2026, ElevenLabs released its v4 speech model, bumping language support from 70 to 90 and — more to the point — closing the gap between a company's primary language and everything else it offers. The same week, the company's valuation reportedly crossed US$22 billion. When a voice-AI company is worth more than a mid-sized Canadian bank, it's probably worth ten minutes understanding what just changed.
What actually shipped
Strip away the launch-day noise and three things matter for anyone who runs a phone line.
First, breadth. Ninety languages, with the company singling out the biggest quality jumps in Japanese, Brazilian Portuguese, Mandarin, and Cantonese. For a country where more than 200 languages are spoken at home — according to Statistics Canada's 2021 Census — that breadth isn't a novelty. It's close to the whole ballgame.
Second, expression. v4 holds a voice's identity across longer passages and reads context as it goes, adjusting tone the way a person naturally would. ElevenLabs also expanded the inline tags that let you steer delivery, and you can now stack them, so a line can move from warm to apologetic to reassuring inside a single sentence. Anyone who has sat through a robotic "I'm... sorry... to... hear... that" knows why this matters more than it sounds.
Third, speed. The companion v4 Turbo model starts generating audio the moment the language model behind it begins forming an answer, instead of waiting for the full response. On a live call, that's the difference between a natural back-and-forth and the awkward half-second lag that makes a caller say "...hello?"
The bilingual problem was always a voice problem
Here's what most of the coverage misses. The logic of a bilingual AI voice agent — one that greets a caller, detects their language, and carries the whole conversation in it — has been sound for a while. What held it back wasn't logic. It was texture.
A voice agent that handles English beautifully but sounds like a tourist reading from a phrasebook the moment it switches to French, Punjabi, or Mandarin is not bilingual in any way that matters to the person on the other end. People hear the seam. And in a lot of Canadian markets, that seam is expensive.
Think about where your calls actually come from. A renovation contractor in Brampton fielding calls in English, Punjabi, and Hindi. A dental clinic in Richmond, B.C., where a large share of patients are more comfortable in Cantonese or Mandarin. A logistics firm in Winnipeg, a boutique hotel in Halifax, a notary in Gatineau whose callers flip between French and English in the same sentence. These aren't edge cases. They're a normal Tuesday.
Until now, serving all of those callers in a voice that sounded native meant hiring for it, and staffing a multilingual front desk across the country is a genuinely hard and genuinely costly thing to do. v4 doesn't make that problem vanish, but it lifts the quality bar high enough that "let the agent take the overflow in whatever language the caller prefers" stops being a compromise and starts being a plan.
Six time zones, one front desk
There's a geography angle here that gets lost in the AI-voice conversation, which tends to be written as if every business sits in one city with one rush hour. Canada doesn't work that way. A customer in St. John's is finishing lunch while Vancouver is still on its first coffee. If you sell nationally — and more small businesses do than used to — your phone has a four-and-a-half-hour window where "someone is always calling" is literally true.
Covering that window with humans means either paying for overnight coverage or accepting that a chunk of callers hit voicemail. Covering it with a voice agent that sounded wooden in anything but English meant you were still only half-covering it. The quieter story in the v4 release is that the always-on front desk and the multilingual front desk have finally merged into the same thing.
Why Canadian businesses specifically should care
Three reasons, and they stack on top of each other.
The first is demographic. Canada's linguistic map isn't a Quebec-versus-everyone-else story. Toronto, Vancouver, Calgary, Ottawa, and Winnipeg all have large communities whose first language is neither English nor French. A phone system that can only truly perform in English is leaving money, and goodwill, on the table in exactly the cities that are growing fastest.
The second is legal, at least in one province. In Quebec, Bill 96 turned serving customers in French into a firmer obligation rather than a courtesy. A business there that answers English-first and French-maybe isn't only risking a lost sale, it's out of step with the law. An AI voice agent whose French is genuinely good turns a compliance headache into a solved line item. If you want the deeper read on what these agents can and can't do today, we broke that down in this piece.
The third is competitive, and it's the one owners feel in their gut. Your competitor's phone still rolls to voicemail at 6:40 p.m. Yours doesn't have to — and now it doesn't have to pick a single language to be good at, either.
The honest caveats
I'd be doing you a disservice if I pretended this was pure upside.
Better-sounding voices sharpen an old question: should your AI tell callers it isn't human? When the voice was obviously synthetic, the disclosure was baked in — nobody mistook it for the person who usually answers the desk. The more natural these models get, the more transparency becomes a deliberate choice you have to make rather than a side effect of the technology. We've argued before that disclosure is the right call, and v4 makes that stance more important, not less.
There's also the matter of expectations. A voice that sounds human makes callers talk to it like it's human — long, winding, full of "actually, wait, scratch that." The voice can keep up. The logic behind it has to as well. A gorgeous voice bolted onto a brittle script is worse than an obviously robotic one, because it promises more than it delivers and the letdown lands harder.
And the quality jump isn't uniform. The gains are largest in the major world languages. Smaller languages and specific regional accents still vary. Test with your callers, in your languages, before you trust it with your busiest line. A demo recorded in a studio tells you what the model can do; a recording of your own after-hours calls tells you what it will do for you.
The great equalizer for small shops
There's a quieter consequence worth naming. For a long time, truly multilingual phone service was a big-company luxury. A national telecom or a bank could staff a French line, a Mandarin line, a Punjabi line, and route callers accordingly, because the volume justified the payroll. The three-person flooring company across town could not. It answered in English, did its best when someone asked for French, and hoped the caller was patient.
That gap was never about ambition. It was about math. You can't justify a dedicated Cantonese-speaking receptionist for the four or five calls a week that need one, even when losing those four or five calls stings. So smaller businesses quietly accepted a thinner version of service than their larger competitors, and customers quietly learned which door to knock on when they wanted to be understood.
A capable bilingual — or trilingual, or more — AI voice agent collapses that gap. The flooring company and the national chain now draw from the same quality of voice. The small shop no longer needs the volume to justify the coverage, because the coverage stops scaling with headcount. If you've ever felt that your size forced you to compete with one hand tied behind your back on service, this is the release valve. It won't win the job on its own. It stops you from losing it before you've said hello.
What I'd actually do with this
If you run a business where calls come in more than one language — and in Canada, that describes most of you — the move isn't to rip anything out. It's to run a real test, this quarter.
Pick your two or three most common caller languages. Have a modern AI voice agent handle after-hours and overflow calls in each one. Then listen to the recordings the way a customer would, not the way a vendor demos them. Ask the uncomfortable question: would a native speaker relax on this call, or tense up?
A year ago, the answer was almost always "tense up" the moment you left English. That's no longer a safe assumption, and assumptions that quietly stop being true are exactly how businesses get out-competed. Once the cost of a genuinely bilingual front line drops from "hire several people across several time zones" to "configure one agent," the arithmetic changes for a lot of companies. We laid out what that arithmetic looks like, in plain dollars, over here.
The bilingual AI voice agent isn't a someday technology anymore. As of late September, it's a this-quarter question. The businesses that treat it that way will be the ones answering every call, in every language a caller brings them, while everyone else is still apologizing for the beep.