Multilingual AI Voice Assistants and Newcomer Service in Canada

Canada has an unusually clear answer to the question "which languages must a service be offered in?" The answer is two: English and French. Not three, not twelve. Two.

That answer was written into the Official Languages Act, and it is a strong framework — equal quality, an active offer, and real remedies when it is not met. The difficulty is demographic rather than legal. Canada's population now includes 4.6 million people who speak predominantly a language other than English or French at home. That is 12.7% of the country, up from 7.7% in 1991, and it is the highest proportion on record.

This article is about what multilingual voice AI can and cannot honestly do about that gap. The short version: it can help, it will not help equally in every language, and the honest failure mode is specific enough to plan around.

Canada's language reality

Statistics Canada's 2021 Census gives a clearer picture than the rhetoric usually does.

• 4.6 million Canadians (12.7%) speak predominantly a non-official language at home — up 16% since 2016, and up from 7.7% in 1991.
• 9 million people — one in four — have a mother tongue other than English or French, the highest proportion since the question was first asked in 1901.
• The two largest non-official languages spoken predominantly at home are Mandarin (531,000) and Punjabi (520,000). Punjabi grew 49% in five years.
• Compare that with earlier waves: Italian, Polish and Greek speakers actually declined, largely because those speakers are ageing.
• More than 180,000 people speak an Indigenous language at home at least regularly — languages that general-purpose commercial tools are not built for.
• 98% of Canadians know English or French, and about 10.7 million — nearly three in ten — can hold a conversation in a non-official language.

Two things are true at once. Official languages are near-universal as a matter of knowledge. And the everyday linguistic reality of roughly one in eight people is something other than English or French.

The legal gap, stated precisely

It is worth being exact about what the Official Languages Act does, because the precision is where the interesting question lives.

The Office of the Commissioner of Official Languages summarises the general rule plainly: federal institutions must communicate with the public in both official languages — on websites and social media included — with communications published simultaneously in each and "of equal quality in English and French."

The duty attaches to head and central offices, to the National Capital Region, to institutions reporting directly to Parliament, and to offices where there is "significant demand." An "active offer" is required: the institution must make it known that service is available in either language, and if it interacts with you without offering that choice, it is not meeting its obligation. Crucially, the duty flows through to third parties acting on the institution's behalf — if a contractor would have to provide bilingual service when the institution does, so does the contractor.

There is also a principle worth pausing on. To assess whether service is substantively equal, the Commissioner looks at equal status, equal use, equal access and equal quality. One illustration: "the wait time for service must not be longer in one language than in the other."

That is a high standard. And it applies to exactly two languages.

The Act creates no obligation for any other language. There is no duty to serve Punjabi, Mandarin, Tagalog, Arabic, Farsi, Hindi or Ukrainian. Interpretation for other languages exists across government, but as a discretionary service decision rather than an enforceable right.

This matters for how the conversation is framed. It is not accurate to say a government is breaching the Act by not serving a third language — there is no such requirement to breach. The honest description is a mismatch between a two-language entitlement and a population of which 12.7% lives in a third. That is a policy question. It is also, notably, a gap that the bilingual obligation itself demonstrates government is capable of taking seriously when it chooses to be.

What newcomers actually experience

The best evidence here comes from the Canada Revenue Agency — which commissioned an ethnographic study of vulnerable newcomers and documented, in its own words, what happens when someone arrives without an official language.

The findings are unusually specific.

• Of 48 newcomers interviewed, only 15 (31%) could complete the interview in an official language. The other 33 (69%) required an interpreter.
• Among refugees resettled between 2010 and 2014, only 26% of government-assisted refugees knew at least one official language, compared with 38% of privately sponsored refugees.
• A 2016 IRCC report on Syrian resettlement found 83% of adult government-assisted refugees self-reported speaking neither official language, with an average Canadian Language Benchmark of 2.6 — very low.
• Free language classes exist, but waiting lists run "in some cases several months long," particularly at lower levels.
• Only one of the 48 had ever personally called the CRA. Newcomers routed around the agency instead, taking questions to settlement workers, private sponsors, friends and family.
• Participants reported that telephone is harder than in-person — it removes non-verbal cues — and that low computer literacy made the online channel inaccessible. Only three of 48 had ever used the CRA website.
• The agency's own conclusion recommended "multilingual documentation and plain language resources" and "minimizing wait times."

The pattern is not primarily a language problem. It is a trust and channel problem that language sits inside. People with the highest need deliberately avoid the institution's own channels, which means a new phone line may not reach the population it was built for — however good its Mandarin is.

Separately, IRCC's 2023 Settlement Outcomes Report found that 32% of newcomers were unaware of the free settlement services available to them at all. You cannot use a channel you do not know exists.

Meanwhile, the phone channel is failing in the languages it does offer

This is the part that makes the multilingual case interesting rather than merely nice.

In October 2025, the Auditor General of Canada reported on the Canada Revenue Agency's eight contact centres. The findings:

• The agency received more than 32 million calls in 2024-25. Just over 10 million reached an agent.
• Its service standard was to answer 65% of calls within 15 minutes. It achieved 18%. In June 2025, 5%.
• Callers who did reach an agent waited an average of about 31 minutes — roughly twice the previous year.
• 8.6 million calls were deflected, meaning the caller was never given the option of speaking to an agent at all.
• About 12% of callers waited more than 60 minutes.
• The Auditor General placed 167 test calls. Responses to general individual-tax questions were accurate 17% of the time.
• Complaints about contact centre service rose 145% since 2021-22.
• Agent headcount fell from 7,782 in 2022-23 to 4,547 in 2024-25, while call volumes rose 30%.

Immigration, Refugees and Citizenship Canada's support centre shows a similar shape. It received 10.64 million enquiries in 2022-23, up 128% from 4.66 million in 2018-19, and answered 4.11 million — about 39%. Its call answer rate fell as low as 13.5%. A published service standard aims for half of calls requesting an agent to reach the queue; in 2022-23 only 15% did.

A 2018 survey found Canadians expect to wait about seven minutes to reach the Government of Canada.

One caveat matters here, and it is not a small one. These are national aggregates for very large federal operations. They describe the shape of the problem at scale. They are not the reader's own metrics, and no small organisation should read them as a prediction of its own phone system. The relevant observation is narrower and more useful: a phone-first service model was already under strain before any multilingual requirement was added to it. Adding languages to a saturated channel does not fix the channel.

Where machine accuracy actually fails

Here is where honesty is required, because this industry routinely overstates multilingual capability.

Peer-reviewed research on automatic speech recognition — the transcription layer inside every voice assistant — finds systematic performance differences that track language and accent.

Other audits found word error rates of 66% for non-native Arabic on a widely used model, 44.2% for Nigerian-accented English in one commercial service, and a 70.6% error rate for female speakers versus 53.5% for male in a Luganda model trained on imbalanced data. Audits across major commercial providers have also found worse performance for Black American speakers, Scottish speakers, and speakers who self-identify as having Indian accents. Age of onset of English is a significant predictor of accuracy — those who learned English later are recognised less reliably.

Now place Canada's languages against those categories.

Mandarin is tonal. It is the largest non-official language spoken at home in Canada. Punjabi, the second, is not.

The implication is uncomfortable and worth stating plainly: an off-the-shelf multilingual voice assistant will tend to perform well for English, French and Spanish speakers, and measurably worse for Mandarin speakers — who are among the people with the least alternative access. Aggregate "languages supported" numbers hide this, because support is not the same as equivalent quality.

There is a further boundary that should not be blurred. Automated interpretation is not appropriate for healthcare, legal advice, immigration advice or emergencies. Those settings need a qualified human interpreter, because the cost of an error is not a frustrated caller but a wrong decision. AI can take a booking, answer a factual question and route a call. It should hand off anything consequential to a person.

The language where accuracy stops being an equity problem and becomes a legal one

Everything above is about fairness. French is the one case where it stops being only fairness and becomes compliance.

The Official Languages Act does not merely require that French be available. It requires service of equal quality in English and French, and the Commissioner reads that substantively — equal status, use, access and quality, with the illustration that "the wait time for service must not be longer in one language than in the other."

Now put a voice assistant inside that obligation. If it transcribes English at one error rate and French at a noticeably worse one, and if a French caller therefore waits longer, repeats themselves more, or reaches a human more often, the inequality is not a rough edge to be tuned later. It is the specific thing the Act exists to prevent. An English-first assistant with French bolted on is not a bilingual service with a limitation; it is a bilingual service with a defect.

Three requirements follow that a "supports French" line on a vendor page does not satisfy:

• Test English and French as separate systems, and compare them. The question is not "does it work in French?" but "does it work as well in French?" Report the gap, per metric, before launch — not after a complaint.
• Make the machine deliver the active offer, not just respond to it. The institution must make it known that service is available in either language. A system that silently detects a language and answers in it has not offered a choice, and a caller who was never offered the choice has not been served to standard.
• Do not treat machine translation as discharge of the duty. The standard is equal quality of service, not the presence of a French version. Translating an English interaction into French is not the same as serving a French speaker, and the difference is exactly what the Commissioner measures.

The asymmetry is worth naming plainly. Mandarin is the harder technical problem, and no one is legally entitled to it. French is the more tractable technical problem, and everyone is. It would be a strange failure to underinvest in the mandatory language and overinvest in the discretionary one — but it is the failure a supported-language list invites, because a list does not distinguish between "we can do it" and "we are equally good at it."

The design constraint most deployments miss

Return to the CRA finding: one of 48 newcomers had ever called.

If the population you are trying to serve does not use the channel you are building, improving that channel does not reach them. It improves service for the people already calling — which is real value, but it is not the same as widening access.

The practical consequences are concrete:

• Intermediaries are the actual channel. Settlement workers, sponsors and community organisations are where these conversations happen. Systems designed with them in mind — information a support worker can relay, references a newcomer can carry — are more likely to reach people than a phone number.
• Awareness is a prerequisite. With 32% of newcomers unaware free settlement services exist, discoverability may matter more than conversational quality.
• Trust is the gate. Telephone removes non-verbal cues, and that is exactly what people with lower confidence rely on. A machine voice removes more.
• Human handoff is not a failure state. It is the mechanism that makes a limited system safe.

Four decisions that have to be made before it answers a call

Most of what goes wrong in a multilingual public deployment is decided before any technology is chosen. None of these four decisions is technical, and none of them can be made by a vendor.

1. Scope — deliberately small. A credible first deployment is one organisation, one phone number, one case system, two or three languages, and two or three low-risk service types. For most agencies that means routine information, a simple service request such as a pothole or a missed collection, and status enquiries. It also means writing down, in advance, what the system must never handle: benefit or program eligibility, immigration matters, enforcement, legal advice, medical triage, payments, and identity recovery. A scope document that lists only what the system does is not a scope.

2. Identity — before any personal record is discussed. The system needs a defined answer to "who am I speaking to, and what may I say to them?" That is harder in a multilingual deployment than it looks, and the difficulty is linguistic rather than procedural. Names, addresses and dates of birth are exactly the fields that transcribe worst across scripts, transliterations and accents — which is to say, in the same languages where recognition is already weakest. An identity check that depends on the machine correctly hearing a Mandarin surname and a street address is a check that fails in the language it most needs to work in. The honest design acknowledges this and falls back to human verification rather than assuming the transcription was right.

3. Accessibility — who a phone line excludes. A voice channel is not automatically more accessible than a web form; it is differently accessible. A phone-only service is difficult or impossible for residents who are deaf or hard of hearing, who have speech disabilities, who have no private place to speak, who share a device, who lack stable connectivity, or who cannot safely say their personal details aloud. Public-sector accessibility obligations generally do not accept "call us instead" as the whole answer. The deployment needs equivalent non-voice routes — web, SMS, keypad — and a human route, offered rather than buried.

4. Ownership when it fails. Every integration fails eventually: an API times out, a request is submitted twice, a call drops mid-way, a caller is routed to the wrong queue. Someone has to own each of those, with a process for finding and correcting the affected cases. A system without a named owner for its failures is a system that quietly accumulates them.

What an honest deployment looks like

If you are considering a multilingual voice assistant for newcomer-facing service, these are the checks that separate a real deployment from a demonstration.

1. Test each language separately, before launch, with real speakers. Do not accept a supported-language list as evidence of equivalent quality. Test the accents you will actually encounter, not the textbook pronunciation.
2. Test English and French for parity, not just presence. The bilingual obligation is one of equal quality, and it is the one part of this that is a legal requirement rather than a service choice.
3. Instrument the handoff. Measure how often callers reach a human, and how long that takes, per language. If one language hands off far more often, that is your accuracy problem surfacing.
4. Disclose the machine. Tell callers they are speaking to an automated assistant and how to reach a person — in both official languages. This is both an ethical baseline and, in many jurisdictions, a legal one.
5. Configure the escalation path deliberately. Anything consequential — health, legal, immigration status, money — should reach a qualified human, with a qualified interpreter where needed. And a caller who asks for a person gets one, without having to argue for it.
6. Give the caller an explicit way to say "I did not understand." Recovery is the feature. A system that cannot be told it has misunderstood will fail quietly, and repeatedly.
7. Preserve what the caller actually said, in the language they said it. Do not hand a human reviewer only a machine translation. The original audio or transcript is the record, and a lossy English rendering of a Mandarin conversation is not a substitute for it.
8. Offer real alternatives, not just more languages on the same line. Keypad, web, SMS, a qualified interpreter and a human route. A phone line supporting twelve languages still excludes anyone who cannot use a phone.
9. Handle the recordings properly. Voice data is personal information under PIPEDA federally, and under BC's PIPA, Quebec's Law 25, and Ontario's PHIPA in those jurisdictions. Consent, retention and access requests all apply. If you cannot answer where recordings are stored and for how long, you are not ready to deploy.
10. Red-team it before launch. Prompt injection, impersonation, attempts to extract another person's record, and requests the system should refuse. A public service is a target, not a demonstration.
11. Measure access per language, not in aggregate. A single "calls handled" figure hides exactly the inequality the deployment was meant to address. Track misunderstanding and repeat-contact rates per language — those are the honest signals.
12. Budget each language as a recurring cost, not a feature. Every language needs its own testing, tuning, content and support, for as long as it is offered. That is why languages should be added after measured performance, never before it.
13. Keep the legal question separate from the service question. Your obligations may be English and French. Your population may not be. Deciding to serve a third language is a service choice, and it should be made deliberately rather than inherited by accident from a vendor's feature list.

What this does not solve

Multilingual voice AI is not a remedy for a policy mismatch about which languages a state owes its residents. It does not remove the need for qualified human interpretation where decisions have consequences. It does not make unfamiliar systems familiar to someone who has just arrived, and it does not replace the settlement worker who is trusted precisely because they are a person.

There is also a risk worth naming. If an automated multilingual channel is introduced while the legal obligation remains bilingual, the predictable outcome is a two-tier service — a fast, cheap, automated channel for people whose languages the system handles well, and a human channel that is still English and French. That is better than nothing, and it is not equity.

The version worth building treats multilingual access as a service-quality problem with measurable per-language results, not as a feature to be advertised. It is best described as a multilingual, human-supervised intake and service-navigation layer — a description that promises three things it can demonstrate (language access, navigation, and connection to an approved workflow) and promises nothing about autonomous decision-making. That is less impressive at launch and considerably more useful afterwards.

If you want to see how the calls themselves sound, you can try a live voice assistant demo, and for the privacy side of recording calls in Canada, our guide to AI voice assistants and Canadian privacy covers PIPEDA, the CRTC rules and recording consent.