MAI-Voice-2.1: 14 serving regions, public preview, no SLA

Is MAI-Voice-2.1 ready for production? Not by Microsoft's own wording: the Microsoft Learn page says the feature is in public preview, provided without a service-level agreement, and not recommended for production workloads. Azure serves it from 14 regions, and the page says you can access it globally.

That is a launch-planning fact, not a quality verdict. Below is what the page says, and a checklist for deciding whether a preview voice belongs in your pipeline yet.

What the Learn page states

The page covers MAI-Voice-2.1 and MAI-Voice-2.1-Flash together. Both use the same Azure Speech API and SDK as other Azure neural and HD voices, so a MAI voice is chosen by name inside an SSML voice element rather than through a new endpoint.

Microsoft's announcement lists $22 per 1M characters for MAI-Voice-2.1 and $15 per 1M characters for Flash on the Microsoft AI announcement. The Learn page itself sends you to the Azure Speech pricing page rather than printing a price.

A pre-launch checklist

Treat each line as a yes or no before real traffic hits a preview voice.

• Contract: does your customer agreement require an SLA for the voice step? A preview has none, so the answer decides everything else.
• Region: the page says you can access the models globally and that Azure routes requests to the 14 listed regions, but it still asks for a Foundry resource for Speech plus its key and region. Confirm data-residency needs against the page before you pick a resource region.
• Fallback: can the pipeline switch to another voice if the preview voice errors or changes? Keep the fallback voice id in config, not in code.
• Cloning: instant voice cloning is gated behind a Limited Access review, so a launch that depends on a brand voice has an approval step ahead of it.
• Licensing: the page says Microsoft holds full licensing rights for commercial use. Read it with your legal team before you resell the audio.

Where an async job changes the risk

A real-time agent notices a failing voice within one turn. A batch narration job does not need to: you can regenerate a failed clip later. That makes batch work the safer first use of a preview model.

Sume's own TTS surface is job-based. A submit returns a job, and you read the result by polling or by webhook, as described in Sume jobs and results. Synchronous mode waits at most 30 seconds, and 30 seconds is a wait budget rather than a job duration. Sume TTS 1.0 is listed at $0.0475 per 1,000 characters in the Sume catalog, a figure you can read live from GET /v1/catalog (see the API reference).

None of that makes Sume a substitute for MAI-Voice-2.1; it is a different layer. It does mean a narration pipeline can keep its audio step behind one job interface, so swapping engines does not rewrite the rest. See the Flash latency comparison for the real-time side.

What to re-check after launch day

Previews change. The Learn page notes that Microsoft adds locales and managed voices as they become available, so re-read the page before each release instead of caching the voice list. Record the date you last read it next to your voice ids.