Microsoft accelerates real-time voice transcription
Microsoft adds new voice models to Foundry capable of transcribing while the person speaks. The challenge goes beyond technical prowess: it affects call routing, latency, and reliability of voice agents.
Microsoft unveiled new voice models in early October within Foundry, its platform for models powering applications and agents, with a simple promise that is harder to deliver: transcribing speech live, without waiting for the end of the sentence. The MAI-Transcribe-2-Streaming model produces its first hypotheses within a few hundred milliseconds, while MAI-Voice-2.1 and MAI-Voice-2.1-Flash are designed to generate a more expressive or more responsive synthetic voice. For contact centers and voice assistants, this matters because every saved back-and-forth reduces silences, interruptions, and dropped calls. And in voice, silence quickly becomes costly.
A transcription that starts before the end of the sentence
Microsoft describes MAI-Transcribe-2-Streaming as an incremental transcription model, meaning a system that offers a partial result before the utterance is finished, then refines it word by word. Naomi Moneypenny, senior director of product development at Microsoft Foundry Models, explains that the first hypotheses arrive within the “low hundreds of milliseconds” after receiving the audio. The company also says the model ranked first in accuracy for both partial and final transcriptions in the Artificial Analysis leaderboard. This approach changes the pace of usage: an agent can identify a customer’s request earlier, and an assistant can prepare a response or tool call while the user is still speaking.
In practice, the promise is not limited to better word recognition. It also aims to reduce the perceived delay in voice exchanges, which remains one of the main friction points in conversational interfaces. Microsoft positions MAI-Voice-2.1 as the “expressive fidelity” model, while the Flash version is designed for high-volume use cases where speed and cost are priorities. The choice between the two therefore depends on the type of interaction, not just raw performance metrics.
Why contact centers are watching closely
Will Bodewes, CEO of Phonely, sees Microsoft’s arrival as a strong signal for the market. According to him, “Voice AI is ready for prime time, if Microsoft’s entry into the market is any indication.” He also recalls that early hosted voice APIs could queue a request behind hundreds of others for 30 seconds, which on a phone call mostly felt like a bad joke. Phonely therefore built its own routing infrastructure to find the least loaded endpoints and bring its response times down to 120 milliseconds, according to Bodewes. The key point is not just speed: it is managing the queue, thus the real user experience.
The CEO also stresses another often underestimated aspect: business safeguards must remain in code, not be entrusted to the model. “Every sentence you add to a prompt adds another chance of failure. So we enforce the critical steps in code.” For him, asking an AI to handle identity verification or refunds alone is opening the door to incidents. Microsoft aligns with this view by targeting customer service agents, multilingual assistants, and smoother voice interfaces, but reliability will still depend on the architecture around the model.
The real challenge remains in the middle of the conversation
Amit Kiran Rege, machine learning researcher at Phonely, tempers the enthusiasm. He believes Microsoft has mainly improved the two ends of the voice chain — input and output — while the main difficulty lies in the middle, in what the system must decide during silences. This is where an agent must know if the speaker has finished, hesitates, or is resuming speech. According to Rege, speeding up the pipeline without adjusting turn-taking logic can have the opposite effect: more audible interruptions.
His observation is concrete. Phonely found that halving latency could make these interruptions more noticeable because the system responded too early before the person had truly finished. In other words, faster transcription alone is not enough to make a good voice agent. Quality also depends on timing, context, and the ability to let the exchange breathe. This is often where demos shine and production reminds us of reality.
What Microsoft aims to establish on the market
With MAI-Transcribe-2-Streaming, MAI-Voice-2.1, and MAI-Voice-2.1-Flash, Microsoft is trying to consolidate a complete voice offering within Foundry, its platform for developers and agents. The company targets fairly specific use cases: customer service, multilingual assistants, live captions, and voice-controlled interfaces. The move is strategic because it places Microsoft on a field where competition is no longer just about the quality of generated text but the entire voice pipeline, from decoding to sound output.
For product teams and operators, the message is quite clear: voice becomes usable at a larger scale, but only if recognition, synthesis, latency, and routing are handled together. Otherwise, you get a very polite agent who interrupts with remarkable confidence. That’s not exactly the desired behavior.
By the numbers
- A few hundred milliseconds — delay for first hypotheses, according to Microsoft Foundry, October 2026.
- 1st place — accuracy announced on Artificial Analysis for MAI-Transcribe-2-Streaming.
- 120 milliseconds — response time claimed by Phonely after routing optimization.
- Two thirds — share of cases where Phonely’s agent responded too early, according to Amit Kiran Rege.
- 3 models — MAI-Transcribe-2-Streaming, MAI-Voice-2.1, and MAI-Voice-2.1-Flash.