Foundry voice agents: public preview, models and costs
Microsoft's 24 September 2026 announcement describes voice agents as a public-preview type in Foundry Agent Service. They bring voice into the agent workflow, but the preview is not a free consumer voice app or a guarantee of every model and channel in every Azure region.
Official publisher logo hosted by its publisher. Official source. Independent guide, not affiliated or sponsored. Artwork and trademarks belong to their owners.
What is a voice-native agent?
The announcement describes building, deploying, tracing and evaluating spoken interactions within the same Foundry platform, API and SDK used for other agents.
Voice still needs the right speech model, permissions and application. Speaking an answer does not make it more accurate or authorise actions on a connected service.
Model choice and Voice Live
The blog names GPT Realtime, Azure Realtime, MAI and bring-your-own options. The current Voice Live overview lists models with separate pricing tiers, including Pro, Standard and Lite.
Check the exact combination you deploy. A broad platform list is not proof that every model supports every language, custom voice, avatar or data-processing region.
Languages and interruption handling
The announcement describes more than 80 languages and 140 locales, natural turn-taking and recovery when a speaker interrupts.
Those are platform claims, not a measured Malayalam call. Test local names, mixed-language speech, background noise, accents, pauses and correction of a misunderstood number.
Deployment channels
Microsoft lists web, Teams, Teams Phone and Twilio-based inbound and outbound telephony. A developer can reuse configured tools, knowledge and governance from other Foundry agents.
Phone delivery can involve separate phone numbers, calling charges, regional rules and customer consent. A sample web agent does not prove a production outbound campaign is ready or permitted.
Custom speech, voice and avatars
The announcement describes domain vocabulary, custom voice and photo or video avatars. Each can have separate access conditions, consent and processing requirements.
Do not clone a speaker or animate someone's likeness merely because a platform offers the feature. Obtain the right permission and read the actual feature terms.
Preview versus general availability
The blog calls voice agents public preview. It labels the Foundry Toolkit in Visual Studio Code coming soon. The broader agent service's availability does not make every voice-related feature generally available.
Confirm current service status, region, quotas and support terms before using it for critical customer work. No production-readiness promise is made by this guide.
Costs are a stack
Model inference, audio, hosted compute, tools, avatars, storage and telephony can all contribute. The Voice Live documentation gives model pricing tiers, not one universal total per agent.
No final INR quote, account free allowance or complete project bill was verified. Set budgets and review current regional prices before running calls. A public preview is not automatically free.
A developer test plan
- Choose one harmless spoken task.
- Check the exact model, region and preview status.
- Review prices and set a small budget.
- Keep tools read-only at first.
- Use a test script without customer secrets.
- Test interruptions and corrections.
- Verify names, dates and amounts in the transcript.
- Review latency, quality, traces and usage.
- Add human handoff for uncertain or urgent cases.
- Require separate approval before real calls or account changes.
India and data-residency limits
A deployment region, a model processing zone and a telephony destination are different facts. Read the exact model's regional and data-processing terms instead of assuming an India resource keeps every step in India.
No Azure resource was created, model deployed, voice generated or phone call made here. Malayalam performance, Indian number availability and a final bill remain untested.
What changed since the original post?
We preserve preview timing, separate coming-soon tooling from current services and add model, region, consent and layered-cost checks. Voice capability claims are not presented as a completed deployment.
Frequently asked questions
Is the preview free?
No universal free entitlement was verified.
Does every model support every voice feature?
Check the exact combination and current documentation.
Was a live agent tested?
No.
Sources: Official September Foundry announcement · Current Voice Live overview and model tiers. Checked 7 October 2026.