Microsoft AI released three speech models on October 1 and put a price on all three. Streaming transcription runs $0.54 per hour of audio. Text-to-speech runs $22 per million characters, or $15 for the Flash version. The figures come from Microsoft’s launch post, and Windows Report lists the same prices.
For a voice-support operation, speech has always been the layer easiest to cost out. Microsoft just posted that cost.
What shipped
MAI-Transcribe-2-Streaming returns its first partial transcripts in just over 100 milliseconds, by Microsoft’s account, and Microsoft says words appear twice as fast as the closest competitor. It also says the model ranks first on Artificial Analysis for accuracy on both partial and final transcripts. That ranking is Microsoft’s citation of someone else’s leaderboard, and we did not check it there. The model handles 60 languages with automatic detection. The $0.54 is an introductory rate through the end of 2026, and the standard price was not disclosed.
MAI-Voice-2.1 and MAI-Voice-2.1-Flash cover 23 languages across 26 locales, and one voice keeps its native accent from language to language. Both clone a voice from a few seconds of reference audio, with consent checks built in, per Microsoft. Flash is billed as 55% faster at inference and about 60% cheaper than comparable models, producing 45 seconds of audio in roughly 150 milliseconds end to end. Microsoft does not say which models it is comparing against.
All three are available through Microsoft Foundry, MAI Playground, Vercel and Azure Voice Live. LiveKit is listed as coming.
The unit math
Start with transcription. $0.54 an hour is under a cent a minute. SiliconANGLE reports that Microsoft’s batch model, MAI-Transcribe-2, costs $0.10 an hour. The streaming version costs 5.4 times that. The premium is not for accuracy. It is for latency: words on screen while the caller is still talking, which is what a voice agent needs and a meeting transcript does not.
Now the voices. Assume ordinary speech runs about 1,000 characters a minute. That is our estimate, not Microsoft’s. At that rate $22 per million characters comes to roughly $1.30 per hour of generated speech, and Flash to roughly $0.90.
Put the two ends of a call together, a live transcriber listening and a Flash voice answering, and the speech fees land near $1.45 an hour. That excludes the language model deciding what to say, which is the expensive part and the part Microsoft did not price here. We do not have a per-hour cost for a human agent seat at Microsoft or anywhere else, so we cannot give the ratio. Any support operations lead can run it against their own books in a minute.
What the launch leaves out
It names no customer. It gives no task-completion rate. That matters because the speech layer was never the hard part. When Google’s Gemini 3.8 Live was tested in September, it scored 97.7% on audio reasoning and finished 35.1% of Sierra’s banking calls. Hearing and speaking are close to solved. Getting the call resolved is not, and nothing in Microsoft’s post moves that number.
Who sits under this price line
Tier-1 voice support comes first. We reported in July that Microsoft’s own support organization shrank from 50,000 to 40,000, and that a Forrester analyst expects half these roles gone by 2030. A price list for live speech is the input that forecast needs.
Live captioners and real-time meeting transcription come next. Batch transcription is already $0.10 an hour, and the live version is now $0.54.
Narration work follows: e-learning, phone prompts, localized dubbing. One voice across 23 languages, billed by the character, changes the unit of work from a studio session to a text file. On the voice-cloning side, the Save Our Voices Now letter in August asked Britain to treat a voice as property, citing three seconds of audio. Microsoft’s price sheet is the first time we have seen a major vendor attach a rate card to the same capability.
What to watch
The introductory rate ends December 31, so Q4 is a pilot window. Enterprises will run a few internal lines on it and report back. The real signal arrives in January. If the standard price lands near $0.54, Microsoft is buying share. If it jumps, the intro rate was bait. Separately, watch for the first named customer. Until one appears, the price is a rate card. Once one does, it becomes a line item on someone’s headcount plan.
