Microsoft is quietly testing a new speech-to-speech model called MAI Realtime that could eventually reduce its dependence on OpenAI's real-time audio technology. The model has appeared in a hidden early-access preview within MAI Playground, allowing a small partner group to test two voices, Victoria and Grant, with configurable interruption handling and multilingual conversations. However, Microsoft has not announced the model, published technical specifications, disclosed pricing, or added it to the public MAI Playground catalog, meaning the details remain unconfirmed by the company itself.

What Makes MAI Realtime Different From Microsoft's Current Speech Tools?

Microsoft's existing speech lineup is divided into separate components. MAI-Transcribe-1.5 converts audio to text and supports 43 languages, while MAI-Voice-2 generates speech from text in 15 languages. These tools work sequentially: speech recognition happens first, then text generation, then text-to-speech conversion. MAI Realtime, by contrast, would handle audio as part of the ongoing conversation rather than as a series of handoffs between separate systems.

The key advantage of this unified approach is what engineers call "full duplex" communication, meaning both sides can communicate at once, much like a natural human conversation. The model would theoretically detect when a user interrupts, stop speaking, retain conversational context, and respond without waiting for silence thresholds to trigger a response. This is significantly harder to achieve than it sounds.

Why Is End-of-Turn Detection So Difficult in Voice AI?

According to the leaked details, MAI Realtime uses two approaches to solve the interruption problem. One method, called Switchboard, reportedly uses an MAI-Ears endpointer controlled by inline tokens. The other combines silence-based endpointing with a Whisper semantic endpointer. Neither component name has been independently confirmed by Microsoft, but the design direction reflects a real engineering challenge: determining whether a pause means "I have finished speaking" or "I am thinking."

Microsoft's existing Voice Live API documentation already emphasizes the importance of this capability. The service offers silence-based server voice-activity detection, semantic voice-activity detection, multilingual semantic detection, server-side echo cancellation, and noise suppression. MAI Realtime, if it launches, would reportedly pull this intelligence closer into the model itself rather than requiring developers to assemble multiple separate services.

Where Does MAI Realtime Fit in Microsoft's Current AI Stack?

Microsoft's Voice Live API currently supports several real-time models, including OpenAI's GPT-Realtime, GPT-Realtime Mini, and GPT-Realtime 1.5. The platform also lists azure-realtime and phi4-mm-realtime as supported options, showing that Microsoft has not been standing still with only OpenAI-hosted real-time audio. However, the unresolved question is where MAI Realtime would sit among these offerings: whether it is a new MAI-branded replacement for an existing component, a higher-capability model for Copilot-scale use, or a separate internal endpoint exposed first through the playground.

This naming ambiguity matters significantly for enterprise customers. Azure developers make architecture decisions based on model availability, regional deployment, data processing terms, token pricing, and whether an API is stable enough for production use. A hidden playground listing answers none of these questions. As of August 2, MAI Realtime does not appear in Microsoft's public Voice Live model table, meaning there is no public developer contract for it yet.

How to Evaluate Voice AI Models for Your Organization

• Test Interruption Quality: Evaluate how the model handles overlapping speech, speaker interruptions, and natural pauses in conversation, as polished voice quality is less important than reliable turn-taking in real-world scenarios.
• Assess Latency and Network Conditions: Test the model with your organization's actual endpoints, headsets, contact-center systems, and network conditions, since browser audio pipelines, Bluetooth devices, and microphone placement all influence real-world performance.
• Verify Regional Availability and Pricing: Confirm that the model is available in your required regions, understand token pricing, and ensure the API is stable enough for production workloads before committing to a migration.
• Review Echo Cancellation and Audio Processing: Microsoft warns that server echo cancellation assumes immediate playback of response audio and can degrade if playback is delayed by more than two seconds, so test this with your specific setup.

What Would MAI Realtime Mean for Microsoft's Relationship With OpenAI?

The leaked report frames MAI Realtime as a way for Microsoft to close a dependency on OpenAI's real-time models. That conclusion is directionally sound, but Microsoft's own product record shows a more complicated picture. Microsoft already offers developers the ability to use OpenAI's native audio behavior through Microsoft Foundry and Azure infrastructure, so the company has not been locked into a single vendor. However, developing an in-house alternative would give Microsoft more control over pricing, feature roadmaps, and data handling policies.

The reported language support includes English, German, Spanish, French, Italian, Portuguese, Japanese, Korean, Chinese, Dutch, Hindi, Indonesian, Arabic, Russian, Turkish, Vietnamese, and Thai, suggesting Microsoft is building a genuinely global product. If MAI Realtime reaches production, it could offer enterprises an alternative to OpenAI's real-time models without requiring a complete architecture redesign.

For now, the model remains in hidden preview with no confirmed launch date, pricing, or public availability. Microsoft has not published a model card, disclosed which regions would receive the model first, or explained how existing Voice Live applications would migrate to it. The real test of MAI Realtime will not be how natural the voices sound, but whether the model can handle the messy reality of human conversation: overlapping speech, language switches, background noise, and the unpredictable rhythm of natural dialogue.