Microsoft launches three voice models for conversational agents: what to take from it

On its website, the Microsoft AI division presented three models on 1 October 2026: MAI-Transcribe-2-Streaming, which turns speech into text as the conversation unfolds, and MAI-Voice-2.1 and MAI-Voice-2.1-Flash, which do the reverse by reading a text aloud. The company presents them as tools for building voice agents able to hold a conversation without dead time.

For transcription, Microsoft states that the first provisional words arrive in a little over 100 milliseconds and that the model ranks first for accuracy on the Artificial Analysis benchmark. These are results announced by the vendor itself: they indicate a direction, not a guarantee of performance on a real phone line, with its background noise and accents.

According to Microsoft’s technical documentation, MAI-Transcribe-2-Streaming recognises 60 languages, including French, and detects the language automatically if it is not specified. The text arrives in two stages: a provisional version that corrects itself while the person is speaking, then a final version for each segment. The documentation explicitly cites call centres and voice assistants among the intended uses.

Two technical limits are worth knowing. A session cannot exceed one hour. More importantly, with the Realtime API, the service does not detect the end of speech by itself: it is up to the application to spot when the caller has finished speaking and then ask the service to finalise the transcription as quickly as possible. This setting, invisible to the end user, nonetheless determines the feeling of fluidity or, on the contrary, of interruptions.

The two voice models cover 23 languages according to Microsoft. The list of ready-to-use voices includes four French voices, one of which offers a style suited to call centres. The Flash version is presented as the fastest, designed for real-time voice agents, interactive voice response systems and call centres; the standard version favours naturalness for long content, such as audiobooks.

Both models can reproduce a voice from a short extract of a few seconds (Microsoft recommends 5 to 60 seconds). Microsoft nonetheless reserves this function for access on request, after review, and requires a recording of the consent of the person whose voice is reproduced: only authorised voices may be used in production. For a business tempted to give its manager’s or an employee’s voice to its switchboard, this is a useful reminder: Microsoft imposes this consent for its service and, more broadly, reproducing a person’s voice without their agreement exposes one to legal risks, to be checked with legal counsel.

The three models are in public preview. Microsoft’s documentation specifies that this preview is provided without a service-level commitment and is not recommended for production use. In other words, a switchboard on which customer relations depend should not rest directly on it as long as that status has not changed.

On pricing, Microsoft announces a launch price of $0.54 per hour of audio for transcription, valid until the end of the year. For voices, it lists $22 per million characters for MAI-Voice-2.1 and $15 for the Flash version. These prices are aimed at developers who assemble a service themselves; they say nothing about the final cost of a complete telephone reception service, which also includes telephony, conversational intelligence and operation.

For a company in France, the question of hosting matters as much as speed. The documentation indicates that the voices are served notably from the France Central, West Europe, North Europe and Sweden Central regions. For live transcription, the regions listed are Sweden Central, Central US and South India, a fourth region, East US 2, being announced as coming soon.

Microsoft specifies that these models are accessible from anywhere in the world and that requests are routed to these regions. Before entrusting customer conversations to a tool built on these blocks, it is therefore legitimate to ask your provider where each step is processed: transcription, response generation and voice synthesis do not necessarily take place in the same location.

This announcement changes nothing overnight for a small business, but it provides a reading grid. Faced with an AI telephone reception offer, here are four simple questions to ask: are the models used in preview or in a stable version? In which countries are calls transcribed and processed? Is a cloned voice used, and with what consent? Is the caller told from the start that they are speaking to an AI and that the call is recorded?

For its part, Alyas announces the AI and the recording at the start of the call, transcribes and summarises conversations, transfers the call to the number chosen by the business when necessary, and its hosting is located in the European Union.