What Microsoft announced
The Microsoft AI division published three models from the MAI family on October 1, on its Tech Community blog:
- MAI-Transcribe-2-Streaming: turns speech into text as the audio arrives, in 60 languages and with automatic language detection. It delivers a first version of the text within a few hundred milliseconds and keeps correcting it until the transcript is finalized when the person stops speaking. It is Microsoft's first streaming transcription model.
- MAI-Voice-2.1: a text-to-speech model Microsoft describes as the most expressive it has built, in 23 languages, for cases where voice quality comes first.
- MAI-Voice-2.1-Flash: a version optimized to respond quickly and handle large volumes, with the same languages.
Naomi Moneypenny, head of product development for Microsoft Foundry Models, explained that together they give developers more options to build natural voice experiences, and that transcribing while the person speaks matters when an application needs to act in the moment. The use cases mentioned include customer service agents, multilingual assistants and live captioning.
Speed, pricing and availability
According to Unite.AI, Microsoft says its transcription model ranks first for accuracy on Artificial Analysis's streaming transcription leaderboard, and that the Flash model generates audio with a latency of about 150 milliseconds. SiliconANGLE, for its part, reports an average of about 320 milliseconds for the first text hypothesis and notes that Microsoft does not guarantee that figure in every scenario, because the network and the model generating the response also add time. These are vendor figures, not independent tests.
The published prices are:
- MAI-Transcribe-2-Streaming: $0.54 per hour of audio, a launch rate through the end of 2026.
- MAI-Voice-2.1: $22 per million characters.
- MAI-Voice-2.1-Flash: $15 per million characters.
All three models are on Microsoft Foundry, MAI Playground, Vercel and Azure Voice Live; the voice models are also on OpenRouter. Both voice models can clone a voice from a few seconds of audio, with built-in consent controls, according to Microsoft.
Why it matters: the race for a voice agent that doesn't keep you waiting
An AI voice agent works as a three-step chain: listen (turn speech into text), think (a language model decides what to answer) and speak (turn the answer into speech). Each step adds delay, and on a phone call a few seconds of silence are enough for the customer to feel they are talking to a clumsy machine. Unite.AI sums up Microsoft's strategy as “clawing back time at both ends” of the loop: transcribing before the person finishes and answering with a fast voice.
SiliconANGLE notes that by separating the listening, reasoning and speaking models, developers can control quality, delay and cost independently. With this announcement, Microsoft competes more directly with specialized voice providers and agent platforms in the automated phone service business, which has become one of the most visible uses of generative AI in companies.
What it means for businesses in Colombia and Latin America
The arrival of giants like Microsoft in real-time voice has practical effects for any company that serves customers by phone:
- More competition, better prices. When the big players publish rates per hour of audio or per million characters, the cost of automating calls tends to fall for everyone.
- More natural conversations. Less delay means fewer awkward silences and fewer customers hanging up before being served.
- Check the language and accent. The sources consulted mention 60 languages for transcription and 23 for voice, but they don't detail performance in Colombian Spanish. Before choosing an engine, test it with real calls from your customers.
- The model is not the whole solution. To answer a call you need the number, the phone system, the rules for handing off to a human advisor, the connection to your CRM and recording with notice to the customer.
- Be careful with cloned voices. A tool's ability to clone voices doesn't mean it can be done without authorization; transparency with the customer remains key.
At 19 80 TIC Cloud we integrate AI voice agents with your cloud PBX and call center, so the voice technology you choose answers on your own business numbers.
