Talking to an AI agent still feels slightly off for one reason: the pauses. The system has to hear you, understand you and answer, and every delay in that chain breaks the rhythm of the conversation. On October 1, 2026, Microsoft AI released three models aimed squarely at that problem, covering both ends of the pipeline: listening and speaking.
Three models, one pipeline
According to Microsoft, all three models were available on the day they were announced:
- MAI-Transcribe-2-Streaming turns speech into text in real time.
- MAI-Voice-2.1 is the company’s main text-to-speech model.
- MAI-Voice-2.1-Flash is a faster variant of the voice model for high-volume, latency-sensitive uses.
Microsoft already had a batch transcription model, MAI-Transcribe-2, which takes a finished recording and returns text. The streaming version is the company’s first real-time transcription model, and it solves a different problem: producing text while the person is still talking.
MAI-Transcribe-2-Streaming: listening in real time
The model accepts audio as a continuous stream over a WebSocket connection and returns text incrementally. Partial results update the current line as the speaker continues, and final results confirm a finished segment. That is what makes live captions, dictation and voice agents feel responsive.
Microsoft’s main claims:
- Languages: transcription in 60 languages, with continuous language detection.
- Accuracy: first place on the Artificial Analysis ranking for both final and partial transcripts, according to Microsoft.
- Speed: a final transcript ready roughly 0.13 seconds after speech ends, according to published comparisons of the launch.
- Error rate: 2.5% word error rate in the figures reported at launch, against 2.73% for xAI’s Grok Voice Transcribe 2.0 in the same comparison.
- Price: $0.54 per hour of audio.
For comparison, the figures published around the launch put Microsoft’s older batch model at $0.10 per hour. The extra cost buys immediacy, which is the whole point of streaming.
MAI-Voice-2.1: speaking in many languages
MAI-Voice-2.1 is described by Microsoft as its strongest multilingual text-to-speech model. It supports 23 languages and 26 locales, and its headline feature is a single voice that carries across all of them with a native accent. A voice agent can switch languages mid-conversation and still sound like the same speaker.
Both voice models can also clone a voice across supported languages from a few seconds of reference audio. Microsoft says consent checks are built in to reduce the risk of misuse. How strong those safeguards are in practice is something the announcement alone can’t show.
MAI-Voice-2.1-Flash: built for speed
The Flash variant supports the same languages and cross-language voices, but it is tuned for workloads where every millisecond counts. Microsoft states a latency of 150 ms end to end for generating 45 seconds of audio, and published comparisons say its inference is about 55% faster than the standard model.
Flash is also cheaper. It is priced at $15 per million characters of generated speech, against about $22 for the standard MAI-Voice-2.1, according to published pricing at launch.
Designed to work together
The pitch is not any single model but the combination. Microsoft says MAI-Transcribe-2-Streaming and MAI-Voice-2.1-Flash are meant to be used together so conversations keep moving naturally, in settings like customer service agents, multilingual assistants and interactive learning or media. A demo called Chatter in the MAI Playground lets people try the pairing.
Where to use them
Beyond Microsoft’s own platforms, the models are available through Vercel’s AI Gateway, which bills at Microsoft’s listed rates without a markup. Developers can call the voice models through the AI SDK’s speech function and the transcription model through its streaming transcription function. The Flash voice model is also listed on OpenRouter.
What to keep in mind
A few caveats apply before treating these numbers as settled:
- Most figures come from Microsoft or its launch coverage. The Artificial Analysis ranking is an external benchmark, but the company is the one reporting its result, and independent tests will tell more.
- Latency depends on the setup. A 150 ms figure for a specific output length does not guarantee the same experience over a real network, in a real application, with real traffic.
- Cloned voices raise their own questions. Consent checks are a stated safeguard, not proof that misuse is prevented.
- Prices are launch prices. Teams comparing providers should check current rates before committing.
Why it matters
Voice has become a main battleground for AI products because it is where users notice weaknesses immediately. A model can be brilliant at reasoning and still feel clumsy if it takes a beat too long to answer or mishears a name. By releasing a transcription model, a standard voice and a fast voice together, Microsoft is offering developers a complete set of parts rather than one component to assemble with others.
Whether that wins developers over will depend less on launch-day claims than on how the models behave in production: how accurate they stay with accents and noisy audio, how consistent the voice stays over long conversations, and how the costs add up at scale.
