Back

Microsoft previews its first AI model built to hear and answer speech live, from a cent a minute

Illustration: sound waves flowing from a desk microphone to a small speaker
Illustration: FrameSignal (AI-generated)

Microsoft has introduced MAI Realtime 1, an AI model that listens to people speak and talks back in real time, for businesses building voice helpers such as customer service lines. It starts at $0.01 per minute of audio, plus a separate charge when it reasons through harder requests, and Microsoft says it beats OpenAI's GPT Realtime 2.1 on quality, speed and price, though its own figures show it behind several other rivals at finishing tasks. It is in a limited preview for now, and Microsoft has not said when it will start charging.

Microsoft has introduced MAI Realtime 1, its first speech-to-speech model: an AI that listens to a person talking and answers out loud as the conversation goes. Microsoft announced it on Nov. 17 at Ignite, its yearly technology conference, alongside other new models for coding, security and images. It is built for companies making voice agents, AI helpers that can carry out tasks on their own, such as answering a customer service line.

Microsoft says the model can handle being interrupted, follow quick replies such as "uh-huh", and wait for a speaker to finish. When a request needs more thought, it draws on a more powerful part of the system that works through the problem step by step before answering, which Microsoft says avoids awkward pauses. Microsoft says it can also pick out a request through background noise or other people talking, and connect to a company's own systems to get things done.

Use of the model starts at $0.01 for each minute of audio. The step-by-step reasoning is billed separately, at $1 for every million tokens it reads and $5 for every million it writes. Tokens are the small chunks of text, often parts of words, that AI models read and write. Microsoft says it beats OpenAI's GPT Realtime 2.1 on quality, speed and price, on public tests and in Microsoft's own tests for its customer service software.

On a test of spoken reasoning questions, Microsoft says its model got 98.9% right, at a cost of about $1.35 per hour of input audio in that test. ElevenLabs' best model scored 98.8% at $4.80 an hour. Google's best scored 97.7% at $3.50 an hour, and xAI's Grok Voice 2.0 scored 97.2% at $4.80. Microsoft's score comes from its own testing, while the rivals' scores come from outside tests. It does less well on a test of whether it can complete the tasks a caller asks for during a spoken conversation. There it completed 57.5% of the tasks, behind several rivals, including xAI's Grok Voice 3.0 at 71.6% and OpenAI's GPT-Live-1 at 74.5%, both of which cost more per hour ($4.80 and $5.83). It still finished more tasks than OpenAI's GPT Realtime 2.1, at 45.7%. Microsoft notes that its results come from specific test setups and that comparisons vary by use.

For now, developers can ask to join a limited preview in Microsoft Foundry, its service for building with AI models, and Microsoft says preview customers won't be billed until a public preview begins. Microsoft says the model works with the same connection developers use for OpenAI's realtime voice models, so existing voice apps can switch to it easily. People over 18 with a Microsoft account can try conversations of up to two minutes with it on Microsoft's MAI Labs test site, for personal use only and not for work or school, in countries where the site is offered. Microsoft asks testers not to share anyone else's voice without permission and to leave out anything sensitive, and warns that the version on the test site may make mistakes and may not work the same way as the final model.

Microsoft says companies including Agoda and Reservations.ai are testing it, and it plans to add the model to its customer service software and to Dragon Copilot, for AI-simulated patient conversations that help nurses practice. Microsoft has not said when the public preview starts or when billing begins.

More on Microsoft