Blog Post

Microsoft Foundry Blog
11 MIN READ

Voice agents in Microsoft Foundry Agent Service now generally available

DONGLI's avatar
DONGLI
Icon for Microsoft rankMicrosoft
Nov 18, 2026

Plus major updates to Speech in Microsoft Foundry

People want to talk to AI as naturally as they talk to each other, with instant responses in their preferred language and voices that sounds Human-like. For organizations, this opens a new way to serve customers, with conversations that feel personal at any scale. Voice agents in Microsoft Foundry Agent Service, now generally available, give teams one enterprise-grade platform to build, deploy, operate and continuously improve those experiences. 

Over the past year, Microsoft has made significant strides in bringing this vision to life. At Ignite 2025, Voice Live API became generally available, giving developers a unified API for natural, real-time voice conversations with any AI agent. At Build 2026, we introduced Voice Live integration with Microsoft Foundry prompt agents and hosted agents, making it easier to build voice-enabled AI experiences. 

Today’s release brings together advanced speech capabilities and comprehensive agent lifecycle management in a unified, enterprise-grade platform. Developers can now build, deploy, operate, and continuously optimize intelligent, speech-to-speech conversational agents, accelerating the journey from prototype to production. 

We're also announcing significant advancements to Speech in Microsoft Foundry: more accurate speech recognition, more natural and expressive voices, and expanded choices across languages, voices, and avatars. Together, these innovations empower organizations to create richer, more human-like voice experiences for customers around the world. 

Voice agents in Microsoft Foundry Agent Service  

Voice agents in Microsoft Foundry Agent Service are now generally available. It's an enterprise-grade platform for developers to build Speech to speech conversational voice agents by providing the leading Speech capabilities and supporting the full agent's lifecycle — from building and deployment to operations and continuous optimization.   

Loading Medius video…

The general availability of voice agents in Microsoft Foundry Agent Service introduces a comprehensive set of capabilities designed to help developers build more natural, intelligent, and reliable voice experiences, while simplifying development and operations at enterprise scale. 

  1. End-to-end voice agent lifecycle management
    Going beyond model APIs, Voice Agent builds on the Microsoft Foundry agent platform to help developers build, deploy, and operate enterprise-ready voice agents through a unified platform. 
  • Build and connect: Developers can create voice agents in minutes, connect them to enterprise knowledge through Foundry IQ, and integrate with rich tools through server-side MCP and client-side function calling, apply guardrails to help keep conversations safe and aligned with business policies, with private-network support for supported enterprise resources and integrations—helping teams control access to sensitive systems and data. 
  • Observe, evaluate, and improve: Built-in voice-native tracing, monitoring, and evaluation make conversations inspectable and quality measurable. For example, if a customer experiences a slow response during a billing inquiry, developers can inspect the turn-by-turn trace to determine whether the delay came from speech recognition, the model, a billing-system tool call, or speech synthesis—and replay the conversation when transcript and audio capture are enabled.  
  • Connect to existing channels with easy setup: Deploy voice agents on the web using audio or text input / output or connect them to existing telephony and contact-center infrastructure—including Teams Phone, Twilio, and Genesys—without rebuilding your customer experience. Bring AI-powered conversations to the channels customers already use, while managing agents through a unified platform. 
  1. More natural, expressive, and intelligent voice conversations
    Deliver engaging, real-time conversations with greater flexibility in speech models and voices. 
  • Expanded choice of speech-to-speech models: Alongside existing GPT Realtime models, we’re introducing new options for more natural voice conversations, giving enterprises greater flexibility to balance conversational experience, agentic capabilities, and cost. 
    • GPT Live 1: Delivers full-duplex, highly natural, fluid conversations—from follow-up questions to complex, multi-step requests. Beyond delegating to specialized models, it can route tasks to Foundry prompt agents, and hosted agents across agent platforms, bringing their specialized knowledge, tools, and workflows into the conversation without breaking its flow. This lets enterprises combine an engaging voice experience with agents that handle work behind the scenes, for experiences where conversational richness and task completion both matter. 
    • Microsoft Realtime Flash 1: Combines low-cost serving with competitive conversational quality and agentic capabilities across typical voice-agent domains, including customer support, appointment scheduling, and routine service requests. It supports experiences where agents need to understand customer intent, sustain a natural conversation, and use tools to take action—not simply answer questions. Its cost-efficient approach helps enterprises move beyond small pilots toward the economics required for millions of calls per day, keeping operating costs manageable as adoption grows. The model is now in preview with Voice Live Lite tier pricing, offering an affordable path to scaling voice automation. Below is our internal test result on τ³-Voice benchmark. Microsoft Realtime Flash 1 achieved the highest reported average score in internal testing: 70.9 versus GPT-realtime-2.1’s 63.7—a 7.2-point improvement. Gains were strongest in telecom (+16.6 points) and retail (+3.5 points), while airline performance was lower (−6.0 points). Against GPT-realtime-2.1-Mini, the average improved by 39.2 points. 
  • More expressive voices: New HD V3 voices bring natural, real-time speech with greater expressiveness, optimized for customer service scenarios. Expanded language coverage includes English, Chinese, Filipino, Indonesian, Malay, and Indic languages. Developers can also control speaking styles and expressions using simple instruction tags, without learning complex markup languages. 
  • Enhanced conversation quality: Improved noise suppression helps voice agents stay focused on the primary speaker, reducing interference from television audio and background voices. Combined with more accurate turn-taking and end-of-speech detection, azure_semantic_vad v2 helps distinguish genuine interruptions from brief acknowledgments and speech directed at others—reducing unnecessary interruptions and premature responses for smoother, more natural full-duplex conversations than previous version. Below table reports our test results on Full-Duplex-Bench v1.5, a public benchmark that tests how a voice model behaves when someone else talks while it is speaking. Each column is the share of test cases where the model did the right thing (0 to 1). The total score is the plain average of the four columns. As you can see, the total score improved from 46.3% to 72.2%, a 25.9-percentage-point gain, with the largest improvements in handling speech directed at others and background voices.
  1. Customized voice and avatar experiences
    Create differentiated, branded experiences through customizable voices and avatars. Organizations can use their own data to develop custom voices that reflect their brand identity. We're also introducing VASA-2 Photo Avatar, which transforms a single photo into a lifelike, upper-body talking avatar. Building on VASA-1, VASA-2 delivers more natural movements and greater control over facial expressions, enabling richer and more engaging visual conversations. 
  2. Flexible multi-agent orchestration for complex voice agents experiences
    Voice agents can seamlessly collaborate with Foundry prompt agents and hosted agents through flexible multi-agent patterns. Developers can choose the architecture that best fits their applications: 
  • Delegation to sub-agents: Voice agents handle real-time conversations while delegating complex reasoning and specialized tasks to backend agents. This keeps voice interactions responsive and low latency without sacrificing reasoning capabilities. 
  • Inline sub-agents: Developers can orchestrate multiple specialized voice agents in a graph-based workflow, each with its own instructions, tools, and voice. Agents can seamlessly transition between tasks during a conversation, improving accuracy, controllability, and cost efficiency. 
  • Hosted agents as conversation engines: Voice agents manage real-time speech input and output while communicating with hosted agents through bidirectional text streaming. This allows developers to preserve existing agent logic while adding advanced conversational voice capabilities.

    T
    ogether, these patterns make it easier to build sophisticated voice experiences that combine real-time responsiveness with powerful reasoning and enterprise workflows.
  1. Simplified development and enterprise integration
    Accelerate voice agent development with a streamlined developer experience and ready-to-use integrations. 
  • Unified developer tooling: Azure Developer CLI (AZD) and GitHub Copilot support help developers quickly build, test, and deploy voice agents using familiar development workflows. 
  • Ready-to-use templates: More than 40 industry-specific agent templates accelerate development across customer support and other business scenarios. 
  • Native telephony integration: Built-in Session Initiation Protocol (SIP) support enables voice agents to connect directly with existing contact center and telephony infrastructure, making it easier to bring AI-powered conversations into established customer service workflows.

    Since the public preview launch, thousands of customers have adopted voice agents across a wide range of scenarios, including customer service, automotive, education, human resources, personal assistants, public services, and more. 

Hear from customers already building voice agents in Foundry: 

"As a leading airline, Gulf Air is proud to work with Microsoft to advance how we use AI across our business. Through our collaboration, we have migrated and enhanced our existing voice agents in Microsoft Foundry, bringing voice, enterprise knowledge and intelligent tools together in a more integrated experience. This enables us to innovate faster and deliver even better experiences for our teams and customers."

Ahmed AbdulJalil Ebranhim, Manager, Digital Innovation & Process Management, Gulf Air 

 

“The Hera companion is a voice agent built on Microsoft Foundry. What began as a proof of concept is now an operational experience. Anyone can talk to Hera: she listens and answers in real time. And she doesn’t just answer, she acts. Ask for a tour, and she turns the spacecraft to show you. Switch to Italian halfway through, and she keeps going. But a great voice isn’t enough; she has to be right. With voice agents in Foundry, Foundry IQ grounding and safety controls come together in one place, so every answer is grounded in ESA’s own mission knowledge and checked against Microsoft’s safety controls and our own rules." 

Daniele Montemaggiori, IT Cloud Services Architect, European Space Agency (ESA)

 

"We tested multiple AI vendors and voice pipelines, and none delivered the customer experience we wanted. Voice agents in Microsoft Foundry finally made AI calls feel like a conversation instead of a phone tree - callers can interrupt, pause to think and speak naturally, and the agent keeps up. With a regional bank customer, our AI agent now handles 30% of calls from end to end. When a caller needs a person, it hands off to a human expert in Microsoft Teams on the same call, with full context." 

Matt Landis, CEO & Founder, Landis Technologies 

What’s new in Speech in Microsoft Foundry - More accurate, natural speech models  

Beyond powering voice agents described above, customers use Speech in Microsoft Foundry for live captions, meeting and call transcription and translation, voiceover for videos and e-learning, and reading content aloud. At Ignite 2026, we are making it more accurate, more natural and offering more choice across three core capabilities: speech to text, text to speech and avatars, and speech translation. 

Speech to text: understand more languages, faster and more accurately

Speech to text converts streaming and file-based audio into high-quality transcripts for voice agents, live captions, contact center analytics, and meeting transcription. These scenarios increasingly involve many languages, speakers who switch languages mid-conversation, and large volumes of recorded audio. Our latest updates, all now generally available, address these needs across real-time transcription, Fast Transcription, and containers.

  • Real-time transcription now offers native multilingual recognition across 50+ languages, with more on the way, so you no longer have to specify the language up front. Results arrive in about 100 milliseconds, with availability in select regions. Multichannel support transcribes two audio channels, such as agent and customer, separately and detects the language on each. For single-language audio, post-stream refinement now spans all language locales. To learn more, see our latest blogs on multichannel transcription and multilingual transcription. 
  • Fast Transcription and LLM Speech are now faster than ever, processing audio at a 100+ speed factor across 100+ multilingual languages, including expanded Indic language support. Mixed-language audio can now be transcribed in a single call. LLM Speech ranks among the leaders on Hugging Face's Open ASR Leaderboard and holds the top spot on the multilingual leaderboard, underscoring its quality across languages. Read more in our latest blog post. Microsoft AI's MAI-Transcribe-2 is now available as an additional model option. 
  • On-Prem Containers now support Fast Transcription, letting you transcribe audio files on-premises with GPU inference through the same API you use in the cloud. Learn more  
Text to speech and TTS avatar: more natural voices, more choice, more lifelike avatars 

Microsoft text to speech voices and avatars power many scenarios beyond voice agents mentioned in section 1 above. In addition, natural and expressive voices are required for voiceover, e-learning and read-aloud content, and other apps need voices that sound clear and engaging. At Ignite 2026, we are delivering two updates: a better and broader set of Neural HD standard voices, and the new VASA-2 photo avatar. 

  • New update for Neural HD standard voices: Besides the Neural HD v3 voices enhancement covered in voice agent scenario, we also extend our Neural HD standard voice offering with the Partner Voices (Public Preview), which brings HD voices built by partners such as DataoceanAI and Telus into Microsoft Foundry. Partner voices give customers more choices to find the right one for both their voice agents and content creation scenarios. 

Locale 

Voice Name 

Sample Script 

Audio Link 

En-CA 

TTS-354-F1 

Oh no, I completely understand how that must feel—it can be really frustrating when things don’t go the way you expect. But don’t worry, you’re not alone. I’ll be with you every step of the way and make sure we get this sorted out quickly. Let’s take it one step at a time, okay? 

Ko-KR 

TTS-038 

모두 50대 이상 중년층 고객이 가장 많이 증가한 것으로 나타 났다. 

Zh-CN 

TTS-373-F231 

哎呀,我不是故意让你失望的!这样吧:如果你想要古怪又有趣、充满笑声的体验,就去看蓝人乐团。如果你想要令人惊叹又充满魔力的表演,太阳马戏团是最佳选择。两者都会让你的第一次约会难忘,所以无论选择哪一个都不会错。

  • New VASA-2 photo avatar (Public Preview): Avatar brings a face to voice agent and empowers content creation scenarios such as marketing video or training materials. VASA-2 photo avatar is specially designed to enrich the voice agent experience to be more engaged and personal.  In Ignite 2026, we bring VASA-2 turns a single photo into a lifelike, upper-body talking avatar, an important upgrade from the talking head of VASA-1. It also moves more naturally and gives you more control over facial expressions to fit your scenario. Learn more.
Speech translation: talk across languages in your own voice 

Language should not stand between people. Live Interpreter API, our real-time speech-to-speech translation service, now runs on a new LLM-based model that translates speech while people are still talking.  

  • Breakthrough translation quality: 30% fewer translation errors than the previous production model. 
  • Fast response: Translates end to end in under 3 seconds, on par with professional human interpreters. 
  • Your voice in every language: With Personal Voice, Live Interpreter API keeps the speaker's voice, style and tone in each language, so conversations feel natural and personal. Learn more 

Customers are already using Live Interpreter for contact center agent assist, on phones and smart devices such as earbuds and AI glasses, and in international meetings. 

Built on the same technology, AI Interpreter is now expanded to large-scale events in Microsoft Teams Events, making it easier for organizations to connect with global audiences across language barriers. Organizers can enable AI-powered simultaneous interpretation for their events, and attendees can simply choose the language they want to hear, making the experience more inclusive and accessible. 

Get started today 

Build voice agents: Getting started only takes a few steps: 

  1. Open the Microsoft Foundry portal go to Build, Agents and New agent, select “Voice”. 
  2. Choose a voice-optimized model and configure your agent's instructions.
  3. Connect knowledge sources and tools as needed for your scenario.
  4. Test conversations directly in the built-in voice playground.
  5. Deploy to your desired channels and monitor performance using built-in observability and evaluation capabilities. 

Explore Speech in Microsoft Foundry. Use the Speech playgrounds in Microsoft Foundry Portal to try Speech transcription, voices, avatars and Speech translation. When you’re ready to integrate speech into an application, start with the Speech SDK and Speech in Microsoft Foundry documentation. 

Updated Nov 17, 2026
Version 1.0
No CommentsBe the first to comment