Microsoft MAI Audio: What Streaming Adds to Voice Conversations

Microsoft introduced three new audio models on October 1, 2026: MAI-Transcribe-2-Streaming for live speech recognition, plus MAI-Voice-2.1 and MAI-Voice-2.1-Flash for speech generation. Applications can use them to display text while someone is speaking and then produce a spoken response. The practical advance comes from combining these components; an audio model alone does not provide a complete conversational assistant with its own decision logic.

Key takeaways

• MAI-Transcribe-2-Streaming returns provisional and confirmed text segments from a continuous audio stream. Microsoft says it supports 60 languages.
• MAI-Voice-2.1 and its Flash variant generate speech from text and support 23 languages, including German, according to Microsoft.
• Announced pricing is $0.54 per audio hour for streaming transcription through the end of the year, and $22 or $15 per million characters for speech generation.
• Microsoft Learn labels the new Azure features as a public preview without a service-level agreement and currently advises against production workloads.

Text that appears while someone is speaking

A conventional recording can be transcribed after it ends. Waiting that long would be disruptive in a conversation. Streaming speech recognition therefore processes incoming audio continuously and returns intermediate results. This can give someone immediate feedback while dictating or display captions during a lecture. It can also help an assistant recognize earlier what a request is likely to be about.

The word “likely” matters. Microsoft’s technical documentation explicitly distinguishes provisional text from confirmed segments. An intermediate result is replaced when recognition improves with more audio. An interface must therefore avoid simply appending every new piece of text. Otherwise, it can produce repetitions or multiple versions of the same statement.

Users should be able to see this distinction: live text may still change, while completed segments appear stable. A later correction is manageable in a note. If an assistant derives an order or an appointment change from something it heard, the same correction carries different consequences. The technical distinction therefore creates a design task: provisional understanding and a binding action need different thresholds.

The direct Realtime interface also requires work on conversational turn-taking. According to the documentation, this integration currently provides neither automatic server-side detection of the end of speech nor automatic commitment of the audio buffer. The application must trigger commitment itself, for example at a detected natural pause. Fast text output does not, on its own, establish when someone has actually finished explaining a request.

Two voice models for different tasks

For output, Microsoft has divided the work. MAI-Voice-2.1 is positioned for expressive speech generation and longer content. MAI-Voice-2.1-Flash is aimed more at conversations where response time matters and at high request volumes. Both turn supplied text into audio. Choosing the model that writes that text, retrieves information, or processes a task is a separate decision for the application developer.

A recognizable voice across language changes may be useful for a learning application. For a service assistant, short pauses, clear pronunciation, and behavior during interruptions also matter. The same tradeoff between longer narration and fast interaction is relevant to other new voice models, such as Eleven v4. A good audio sample for a long text does not establish that a conversation will flow smoothly.

Microsoft publishes speed figures for its models. These describe specific processing steps and measurement conditions, rather than the entire waiting time of a finished assistant. The path between microphone and speaker also includes the network, transcription, answer generation, and possibly tool calls. A useful measurement for a prototype is therefore how long a person waits after finishing a sentence for an appropriate response, and whether the system lets them finish speaking.

Microsoft also specifies controlled access for custom voices. The Azure documentation requires approval and recorded consent before a personal voice can be created. This is separate from the voices already offered. The ability to reproduce a voice from a short reference clip therefore does not provide unrestricted access to clone anyone’s voice.

How to try the new components

A manageable starting point is an interface for live captions. Developers can focus on audio input, provisional text, and final segments without also building answer logic. Microsoft documents both a direct Realtime connection over WebSocket and the Azure Speech SDK, a library for integrating transcription into custom programs.

According to the documentation, the SDK handles connection management and recovery. The application is still responsible for presentation: keep confirmed text separate from the current intermediate result, display connection errors, and avoid saving unconfirmed text as final after a failure. These are concrete requirements for useful captions, not cosmetic details.

Alternatively, Vercel offers the models through its AI Gateway. The provider demonstrates streaming transcription and spoken response generation with its AI SDK. OpenRouter also lists both voice models. These access routes simplify integration into existing applications, but do not replace checking the relevant endpoint, its availability, and its billing. A model listing does not establish that every specialized feature is available identically everywhere.

The prices use different units: audio hours for recognition and characters for output. Anyone estimating the cost of a conversational assistant must track both sides separately. The model generating answers and external services may add further costs. The introductory transcription price announced through the end of the year is therefore neither a permanent rate nor the total cost of a conversation.

The next step is a measurable conversational prototype

The new models enable a straightforward experiment: transcribe short German requests, process confirmed text, and read an appropriate answer aloud. Examples should include pauses, self-corrections, proper names, and background noise. Those cases reveal whether early partial results help the application and whether subsequent confirmation is handled correctly.

The public preview also sets a clear boundary for use. For a prototype, it is an opportunity to experiment. According to Microsoft’s own documentation, it currently does not constitute a firm production commitment. The relevant advance is therefore a more flexible combination of listening and speaking. Whether it becomes a helpful conversational partner depends on transitions and everyday situations that a model announcement captures only in part.