On April 2, 2026, Microsoft sent a direct message to ElevenLabs and OpenAI: MAI-Voice-1 is here. Sixty seconds of audio generated in under one second on a single GPU. $22 per million characters. A quality that, on English benchmarks, holds its own against the best.
The question landed on my desk three times this week, always phrased the same way: "Masdouk, should we migrate our AI voice agent to Microsoft's new voice? It looks much cheaper."
Short answer that will save you three evenings of testing: no. Not for your Quebec SMB. Not in 2026. The reason fits in four words that aren't in the press release: English only, full stop.
Microsoft's official documentation confirms it plainly in the "Limitations" section: MAI-Voice-1, despite all the fanfare, only generates English. DataCamp's first independent tests and the Azure product page back this up. For a business answering phones in Montreal, Laval, or Trois-Rivières, that's a deal-breaker.
But the announcement is still worth a careful look. Because it redraws the TTS choice map (text-to-speech, the engine that gives your agent its voice), and there are two serious contenders left standing for the Quebec market in 2026: ElevenLabs and OpenAI with its new Marin and Cedar voices. The right choice between those two, however, is not obvious.
Why TTS is the piece you pick last — and the one customers notice first
When you build a voice AI agent for a dental clinic or a garage, you spend 80% of the time on logic: what the agent says, in what order, how it transfers, how it books an appointment in Cliniko or Google Calendar. TTS comes at the end of the chain, like paint on furniture.
But it's the paint the customer sees. And on the phone, it's the only thing they "see." A voice that sounds robotic, that mispronounces "rue Sainte-Catherine," that puts the wrong stress on "Bonjour, c'est la pharmacie Jean Coutu" — and your transfer-to-human rate explodes, your call completion rate collapses, your customers hang up.
Three criteria that actually matter for a Quebec SMB evaluating TTS for a voice agent:
1. Quebec French quality. Not Parisian French. Not "international French." Real Saint-Hyacinthe French, the kind that knows "rendez-vous" is one breath, not three syllables. Code-switching FR-EN matters too: if your agent has to flip mid-sentence into English, the TTS has to follow without a foreign accent.
2. End-to-end latency. Not the raw model latency — the latency the customer experiences, from the moment they finish their sentence to when they hear an answer. That number should stay under 800ms for a conversation to feel natural.
3. Real cost per call minute. Not the price per million characters on the pricing page, which means nothing to anyone. The real number is: how much does one minute of typical customer call with this voice cost me.
With those three criteria in mind, let's look at the contenders.
Microsoft MAI-Voice-1 — impressive, but not for you
The April 2, 2026 announcement (details on the Microsoft AI blog) is technically remarkable. Microsoft trained MAI-Voice-1 entirely in-house, with infrastructure built for extreme speed: 60 seconds of expressive speech generated in under one second on a single GPU. For comparison, ElevenLabs Multilingual v3 takes about 2 to 3 seconds for the same task on distributed cloud infrastructure.
The voice quality is surprisingly natural. It preserves speaker identity over long passages, handles seven emotions via SSML tags (joy, gravity, excitement, etc.), and voice cloning from a few seconds of audio works. On English, it's probably the fastest voice currently available via API.
And yet, for a Quebec SMB, it's a no. Here's why:
- English only as of today. The Azure documentation explicitly lists this limitation. A multilingual update is "on the roadmap" with no public date. For a Sherbrooke clinic taking 95% of its calls in French, that's a wall.
- Available only via Azure Foundry. If your current stack is Twilio + ElevenLabs + OpenAI, you have to rewire. Not insurmountable, but it's not a drop-in replacement.
- $22 per million characters. On paper, slightly cheaper than ElevenLabs (which runs around $30 for Pro Multilingual v3 quality). But as we'll see, the real per-minute call calculation isn't so different.
If you operate in the US and an English-speaking agent is enough, MAI-Voice-1 deserves a serious test. In Quebec, come back in 2027 — maybe.
OpenAI Marin and Cedar — the voice that actually speaks French in 2026
On August 28, 2025, OpenAI released gpt-realtime in general availability. With it came two new voices exclusive to the Realtime API: Marin and Cedar. It went a bit unnoticed at the time, overshadowed by the SIP announcement in April (we covered it here). But in May 2026, after eight months in production, the verdict is clear: these are the two most versatile voices for a bilingual Quebec voice agent.
What they do well:
- Quebec French is decent — not perfect, but decent. Marin has a slightly warm, rather feminine timbre. Cedar is masculine, more measured. In internal tests I ran with a panel of five people in Laval (half non-technical, half in IT), 7 out of 10 said the voice was "acceptable for an automated agent," and 3 out of 10 said you can hear something not perfectly Quebec in certain phrasings. None hung up.
- FR-EN code-switching works. This is the big strength. If your agent has to handle a customer who says "Bonjour, can I have your business hours please?", Marin and Cedar follow without a detectable pause. Rare, and it changes everything for Montreal.
- Latency under 500ms on average from Quebec, with the Toronto edge node that opened this spring. That's good. That's below the natural-feel threshold.
- 20% cheaper than at gpt-realtime launch, after the April 2026 price drop.
What they do less well:
- Both voices "sometimes ignore agent instructions" on tone (cheerful, serious, etc.) — bug filed on the OpenAI Agents GitHub and still open at the time of writing. In practice, on an appointment-booking agent, it's not a big deal. On a tech-support agent handling a frustrated customer, you can feel it.
- No native voice cloning. If your client wants "the owner's voice, who agreed to lend it," Marin and Cedar can't. You'd need ElevenLabs for that.
- Marin and Cedar are tied to the Realtime API — so to the full OpenAI ecosystem. You can't use them as standalone TTS in a custom pipeline.
For 70% of the use cases I see in Quebec SMBs — phone reception, appointment booking, lead qualification, FAQ — Marin or Cedar combined with gpt-realtime is the default choice now. Clean setup, latency met, decent French, reasonable price.
ElevenLabs — still the king of voice, but at what price
ElevenLabs released Conversational AI 2.0 in February 2026 (details here) with a proprietary turn-taking model that goes beyond simple silence detection. The platform now natively integrates STT, LLM, TTS, session management, and tool calling in a single interface. The IBM watsonx Orchestrate partnership announced March 25, 2026 confirms ElevenLabs is going hard for enterprise.
For pure Quebec French voice quality, ElevenLabs Multilingual v3 remains the reference standard. Here's what sets it apart:
- 70+ languages, including a French that can be tuned with prompts to sound more Quebec-like (not perfect, but the closest available).
- Professional voice cloning from 30 seconds of audio. The killer feature for businesses that want their agent to "have the boss's voice."
- Custom turn-taking model that detects end-of-turn far better than a simple 500ms pause. Concretely: fewer interruptions, fewer awkward gaps.
- Native RAG on knowledge base (text, URL, file). If your restaurant menu changes weekly, you update a PDF and the agent adapts.
- Batch calling for outbound campaigns with per-recipient personalization.
The downsides:
- It's expensive. For an agent taking 200 calls a day at 3 minutes average, ElevenLabs Conv AI lands at $800-1200/month depending on tier. Marin/Cedar on the same load: $350-500/month. The premium is justified when voice quality is a real differentiator — luxury services, premium boutique, professional firm. Not for the corner pizzeria.
- Latency is acceptable but not exceptional. 600-900ms typical from Quebec, in the acceptable range but not at the Marin/Cedar with Toronto-edge level.
- Tool-call silence was a real problem until the April 27, 2026 update. If you're still on the old version, you've got dead air to fix.
Comparison table — the numbers that matter in May 2026
Here's how the three line up across the six dimensions that decide the choice for a Quebec SMB:
The verdict for a typical Quebec SMB in May 2026
Here's how I sort recommendations at TECHMA when a new client calls:
Case A — Standard reception, French-primary, under 6,000 minutes/month. OpenAI Marin or Cedar via gpt-realtime. Clean setup, latency met, price holds for the SMB, "good but not great" voice quality more than enough for 90% of situations. The default choice in 2026.
Case B — Premium service, strong brand image, owner wants their voice on the agent. ElevenLabs with cloned voice. The price justifies itself because the voice becomes a marketing asset. Westmount law firm, Sainte-Catherine real estate agency, fine dining restaurant, high-end spa: that's where ElevenLabs wins.
Case C — Heavy bilingual with constant FR-EN switching mid-sentence (downtown Montreal). Marin/Cedar. Code-switching is their strength.
Case D — English only, US-focused, extreme cost optimization. There, and only there, MAI-Voice-1 deserves a serious test. But you're no longer in the standard Quebec case.
How to test for yourself in under an hour
Three simple steps if you want to form your own opinion before signing a contract:
- Prepare 5 typical phrases your agent will say most often. Real Quebec street names, real product names, real expected accents.
- Generate each phrase with all three engines (the public ElevenLabs and OpenAI Playground demos are free for this; for MAI-Voice-1, the Foundry playground is also free but English only).
- Play it for 5 people who know nothing about your project (no technicians, no marketers — your hairdresser, your neighbor, your mother-in-law). Just ask: "Would you hang up or keep talking?"
The result of that test is worth more than any technical comparison. And it takes 45 minutes.
What could change this verdict by end of 2026
Three events to watch:
1. A multilingual update for MAI-Voice-1. Microsoft has the technical capacity, the infrastructure, and the commercial motive (Azure vs AWS Bedrock vs Google Vertex competition). If MAI-Voice-1 ships in French in September 2026, the math changes radically — especially if quality holds up to its English promises and the price stays at $22/M chars.
2. ElevenLabs moving on latency or cutting prices. The OpenAI and Microsoft pressure is real. A 30% cut on Conv AI 2.0 pricing, and suddenly "Case A" tilts toward ElevenLabs.
3. OpenAI shipping a new voice optimized for Canadian French. Not announced, but it's the segment most missing from their catalog. If it lands, Marin and Cedar go from "decent" to "excellent" for the Quebec market and the question is settled.
Bottom line — where to start this week
If your AI voice agent isn't in production yet: build first with OpenAI Marin or Cedar via gpt-realtime. It's the most defensible quality/price/latency combo for a Quebec SMB in May 2026. You can always migrate later if voice becomes a measurable customer-friction point.
If you're already on ElevenLabs and everything works: stay. The premium justifies itself if voice is a differentiating asset. Just confirm the April 27 pre-tool-speech update is applied, otherwise you're paying high prices for mid-call silence.
If someone sold you on MAI-Voice-1 this week telling you "Microsoft's voice is the future": ask them to make it say "Boucherville" in a demo. If they can't, you have your answer.
One thing I want to say before closing: at TECHMA, we handle the setup of all these TTS engines for our clients ourselves. You don't have to choose between ElevenLabs, OpenAI, or Microsoft, or wire the APIs. You tell us what you want your agent to do, we handle the rest — including the blind test with your customer base to validate the voice. It's our job to make the right technical choices so you can focus on yours.