Microsoft MAI Realtime 1 for Product Managers: Enterprise Voice AI After Ignite 2026
TL;DR
Microsoft announced MAI Realtime 1 at Ignite on November 17, 2026: a bidirectional speech-to-speech model for enterprise voice agents, priced at $0.01 per audio minute with reasoning billed separately. It joins MAI-Voice-1 (TTS) and MAI-Transcribe-1 (STT) in a full Microsoft voice stack, is compatible with the OpenAI Realtime API, and is currently in limited-access preview on MAI Labs and Microsoft Foundry. This guide covers what it is, when the speech-to-speech architecture beats a cascaded pipeline, how to run a voice agent pilot, and what AI PM teams should evaluate before committing to the stack.
The AI PM Minute
One tactic to make you a sharper AI PM, twice a week. 60 seconds to read. Free.
No fluff. Unsubscribe anytime.
What MAI Realtime 1 Is and What It Is Not
MAI Realtime 1 is a speech-to-speech model. That distinction matters. It does not convert voice to text, pass text to a language model, then convert the response back to speech. It processes audio in and produces audio out, with reasoning happening inside the model on the audio stream directly.
Microsoft built it for enterprise conversational AI agents: contact center automation, field service voice assistants, and meeting-aware copilots. The design goals are low latency, natural interruption handling, background noise filtering, and enterprise tool integration via connected actions.
What it is
A bidirectional speech-to-speech model for building voice agents. Audio goes in, reasoning and tool calls happen inside the model, audio comes out. Handles interruptions and silence natively.
What it is not
A replacement for MAI-Voice-1 (TTS) or MAI-Transcribe-1 (STT) in document narration, podcast production, or batch transcription pipelines. Those are separate products optimized for different jobs.
Current access
Limited-access preview on MAI Labs and Microsoft Foundry as of November 17, 2026. Not generally available. Apply through the MAI Labs waitlist.
API compatibility
Compatible with the OpenAI Realtime API. If your team already has a prototype on OpenAI's real-time voice endpoint, you can test MAI Realtime 1 against the same interface with minimal code changes.
Speech-to-Speech vs. Cascaded: The Architecture Decision Every Voice PM Faces
Every voice agent team eventually hits the same fork: build a cascaded pipeline (STT to LLM to TTS) or move to a speech-to-speech model. MAI Realtime 1 is Microsoft's answer to the second path. Neither is universally better. The right choice depends on what your product needs most.
Cascaded pipeline: where it still wins
Cascaded gives you independent evaluation at each step. You can benchmark your STT separately, swap your LLM without touching TTS, and log the full text transcript for quality review. If your product lives in a regulated industry where you need audit logs of every spoken utterance, cascaded is the safer architecture today.
Speech-to-speech: where it wins
Latency and naturalness. A cascaded pipeline compounds latency across three round trips. Speech-to-speech cuts that to one. The model also handles prosody, pace, and emotion natively, which a TTS step reading a text string cannot replicate. For consumer-facing voice agents where response speed is the primary UX driver, speech-to-speech wins.
The observability trade-off
Cascaded pipelines produce a text transcript you can read, search, and feed into your existing analytics stack. Speech-to-speech agents produce audio. You need a transcription layer anyway if you want text logs, which partially erases the latency advantage for some architectures.
The evaluation gap
Evaluating speech quality is harder than evaluating text quality. You cannot run your speech-to-speech agent output through a standard LLM judge without an intermediate transcription step. Voice agent evaluation frameworks are still maturing, and your team will spend engineering cycles building them before you can ship confidently.
Microsoft's bet with MAI Realtime 1 is that the enterprise contact center market, where call handle time is the primary metric, cares more about natural conversation speed than transcript auditability. For other use cases, run the architecture decision as an explicit product choice before you build.
Pricing and the Microsoft Voice Stack
MAI Realtime 1 is priced at $0.01 per audio minute for base speech-to-speech usage. Reasoning is billed separately at approximately $1 per million input tokens and $5 per million output tokens. These are preview prices; expect them to change before general availability.
MAI Realtime 1
Speech-to-speech for real-time voice agents
Pricing: $0.01 per audio minute + reasoning costs
Limited preview. Reasoning billed separately at ~$1/M input, $5/M output tokens. Tone tunable via prompts.
MAI-Voice-1
Text-to-speech for narration, brand voice, assistants
Pricing: Per-character pricing (see Azure AI pricing page)
Best for document narration, podcast production, and cases where naturalness beats real-time latency.
MAI-Transcribe-1
Speech-to-text for high-accuracy transcription
Pricing: $0.36 per audio hour (2.5x faster than predecessor)
Pairs with MAI-Voice-1 for a matched Microsoft voice pipeline in cascaded architectures.
The enterprise angle: Microsoft positioned this stack for Azure customers who want a fully integrated voice pipeline without routing audio data through external vendors. If your enterprise has a Business Associate Agreement with Azure and you are building in healthcare or financial services, keeping the voice stack inside Azure removes a separate compliance review for a third-party STT or TTS provider.
Learn to Evaluate AI Infrastructure as a PM
The AI PM Masterclass covers how to make build vs. API decisions, evaluate model capabilities, and design AI products that hold up in production, taught live by a Salesforce Sr. Director PM.
Use Cases: Where MAI Realtime 1 Actually Fits
The preview access criteria gives away the target: eligible use cases are enterprise conversations at scale. Here are the four domains where the design decisions in MAI Realtime 1 map cleanly to real product requirements.
Contact center triage and deflection
High-volume, lower-stakes calls: account lookups, appointment reminders, payment confirmations. The wedge strategy for voice AI is to start here, prove handle time reduction, then expand to more complex calls. MAI Realtime 1 handles background noise and natural interruptions, which are table stakes in contact center environments.
Internal voice assistants for field workers
Technicians, logistics staff, or clinical workers who need hands-free access to information while working. Azure infrastructure means data stays inside the enterprise perimeter. This is the use case where the enterprise compliance argument for the Microsoft stack is strongest.
Meeting-aware copilots
Agents that listen in Teams calls, surface relevant information during the conversation, and take action on connected enterprise tools. Compatible with Microsoft's 350+ Copilot connectors. The real-time nature of MAI Realtime 1 makes it a candidate for agents that respond to what is being said, not just after the call ends.
Voice-enabled product interfaces
Consumer or B2B SaaS products that want to add natural voice interaction. The OpenAI Realtime API compatibility means lower switching costs for teams that prototyped on OpenAI's real-time endpoint. Only viable once MAI Realtime 1 moves beyond limited preview.
How to Evaluate a Voice Agent Before You Ship
Voice agent evaluation is harder than text evaluation because the output is audio. You cannot pipe a voice agent response directly into a text LLM judge. Here is the evaluation structure that works in practice for enterprise voice AI teams.
Task completion rate
Define success for each call type (appointment booked, balance confirmed, ticket created). Measure autonomous completion vs. escalation to human. This is your primary output metric. Report it at P50 and P95 by call type.
Interruption handling
Build a test set of call recordings where the caller interrupts mid-response. Score whether the agent stops speaking, acknowledges the interruption, and handles the new intent. Agents that plow through interruptions fail in production.
Noise robustness
Test against recordings with background noise at three levels: quiet office, busy office, and noisy environment (call center floor, vehicle cab). Score transcription accuracy and task completion under each condition before comparing models.
Latency to first word
Measure the gap between the caller finishing their turn and the agent beginning its response. Under 500ms feels natural. 500ms to 1 second is acceptable. Over 1 second breaks conversational flow. Test this on your actual infrastructure, not in a demo environment.
Safety and policy compliance
Voice agents can say things you did not intend. Build a red-teaming script that covers policy violations, unsupported claims, and sensitive disclosures. Run it before every model update, not just at launch.
Human escalation quality
When the agent cannot complete the task, it must hand off cleanly: tell the caller it is transferring them, provide a summary to the human agent, and not abandon the call. Escalation failure is often worse than a failed autonomous completion.
What to Do Right Now
MAI Realtime 1 is in limited preview. Most teams will not get access immediately. Here is how to use this window productively.
Apply for preview access now
The waitlist will close to new applicants before general availability. Apply through Microsoft Foundry or MAI Labs even if you are not ready to build. Early access gives you time to evaluate without competing with the GA rush.
Prototype the cascaded path first
Build your first voice agent on the cascaded architecture (MAI-Transcribe-1 to your LLM of choice to MAI-Voice-1). This gives you a working baseline with text transcripts you can evaluate using existing tools, before you commit to the speech-to-speech path.
Define your primary evaluation metric
Pick one metric before you evaluate any model: task completion rate for a defined call type. Everything else is secondary. Teams that evaluate on vague criteria like 'sounds natural' never reach a shipping decision.
Build your noise test set
Collect 20 to 30 recordings of your target use case at different noise levels. This set will outlast any single model evaluation and gives you a consistent benchmark as you compare providers.
Check OAI Realtime API compatibility
If you have an existing prototype on OpenAI's real-time voice endpoint, compare MAI Realtime 1 on the same test set once you have preview access. The API compatibility means the switching cost is low enough to run a real comparison.
Build AI Products That Work in Production
The AI PM Masterclass covers voice AI, infrastructure decisions, evaluation frameworks, and the full technical stack AI PMs need to ship with confidence.
Related Articles
Before you go: get the AI PM Minute
One tactic to make you a sharper AI PM, twice a week. 60 seconds to read. Free.
No fluff. Unsubscribe anytime.