Microsoft has introduced MAI-Transcribe-2-Streaming, a low-latency speech-to-text model designed for applications that need words while someone is still speaking.
Launched on October 1, 2026, the model supports 60 languages, continuous automatic language detection and incremental transcripts. Microsoft also reported that it ranked first for both final and partial transcript accuracy in Artificial Analysis’ streaming evaluation.
For marketing and SaaS teams, the important question is not whether the benchmark looks impressive. It is whether a workflow actually benefits from having the transcript in real time. Live captions, voice agents and in-call assistance may. Recorded customer research often will not.
MAI-Transcribe-2-Streaming at a glance
| Dimension | Current position | Why it matters |
|---|---|---|
| Status | Public preview, no SLA, not recommended for production workloads | Best suited to evaluation and controlled pilots today |
| Languages | 60 languages with automatic language detection | Useful for multilingual workflows, but teams still need to test accents and terminology |
| Streaming price | $0.54 per audio hour through the end of 2026 | Live transcription costs substantially more than Microsoft’s batch option |
| Artificial Analysis result | Ranked first for final and partial transcript accuracy at launch | A strong independent benchmark signal, not a guarantee for every language or use case |
| Realtime session limit | One hour | Long calls and continuous applications need session-management planning |
What MAI-Transcribe-2-Streaming changes
Microsoft’s announcement positions MAI-Transcribe-2-Streaming as the real-time counterpart to the recorded-audio MAI-Transcribe-2 model. Instead of waiting for a complete recording, the streaming model produces provisional text while audio is arriving and then refines it as more context becomes available.
Microsoft says the first provisional hypotheses arrive just over 100 milliseconds after receiving audio. It also reports that words appeared twice as quickly as its closest competitor in Microsoft’s own dictation and subtitling tests. That second point should be read as a vendor-reported application test, not evidence that every product built with the model will respond twice as fast.
Transcription is only one part of a live voice system. A voice agent may still need an LLM to interpret the request, a CRM or database lookup, permission checks, a tool call and speech generation before the user hears anything back. We covered that broader response loop in our look at GPT-Live and conversational voice AI.
The accuracy result is strong, but the latency numbers need context
At launch, Microsoft reported that MAI-Transcribe-2-Streaming ranked first on Artificial Analysis’ streaming speech-to-text benchmark for both final and first-partial transcript accuracy. Microsoft highlighted a 2.5% final word error rate, or WER.
WER measures substitutions, insertions and deletions against a reference transcript, with a lower score indicating fewer errors. A strong composite result is useful evidence, but it does not mean a support call in Hindi, a noisy sales conversation or a developer discussion full of product names will reproduce the same score.
The benchmark combines roughly eight hours of streaming audio across AA-AgentTalk, VoxPopuli and Earnings22. There is also an important timing distinction: Artificial Analysis measures first-partial and final latency after detected speech ends, while Microsoft’s just-over-100-millisecond claim begins while audio is still being received. They describe different moments in the transcription process.
For broader context, our reviews of Grok Voice Transcribe 2.0 and Gemini 3.5 Transcribe show why equivalent modes matter when comparing speech models. Batch and streaming benchmarks answer different questions.
The streaming premium only makes sense when latency creates value
Microsoft launched MAI-Transcribe-2-Streaming at an introductory price of $0.54 per audio hour through the end of 2026. Its recorded-audio MAI-Transcribe-2 model is available at $0.10 per audio hour.
At 1,000 hours of audio, that is approximately $540 for streaming versus $100 for batch, before storage, telephony, downstream model calls, analytics, integrations or speech generation are added.
That premium can make sense for live captions, a support assistant preparing relevant information before the caller finishes explaining the problem, or a voice agent that needs to start reasoning during the conversation. A marketing team analysing 200 recorded customer interviews after the calls are finished has a much weaker reason to pay for streaming.
Paying real-time rates for a workflow that runs after the call is buying speed the workflow does not use.
Public preview and workflow limits still matter
Microsoft’s documentation lists MAI-Transcribe-2-Streaming as a public-preview service without a service-level agreement and says it is not recommended for production workloads. That caveat matters more than the leaderboard position for teams considering customer-facing deployments.
Microsoft currently provides two main Azure integration routes. The Realtime API uses a WebSocket connection and an OpenAI Realtime-like protocol, while the Azure Speech SDK provides managed connection handling. Vercel also exposes the model through AI Gateway as microsoft/mai-transcribe-2-streaming.
The Realtime API currently caps sessions at one hour and accepts raw mono PCM16 audio at 16 kHz or 24 kHz. It also does not provide server-side speech detection or automatic commits, so an application needs to decide when a chunk of speech is complete before requesting a final transcript.
That matters because provisional text can change. A user can pause, correct themselves or change direction halfway through a sentence. Triggering a business action from an unstable partial transcript can turn a fast system into a confidently wrong one.
The Speech SDK also has limitations relevant to analytics-heavy applications. Its results do not currently include detected-language information, confidence scores or word-level timestamps. Teams should therefore test the exact integration path they plan to use rather than assuming every capability from Microsoft’s batch model carries into streaming.
Where it could matter for marketing and SaaS
Live assistance
A qualification or support assistant could begin preparing relevant information while a caller is still explaining the problem, reducing dead air and making the interaction feel faster. But actions such as changing an account, sending an email or booking a meeting still need customer context, permissions and confirmation rules. The same principle applies across AI automation in sales.
Teams comparing complete voice-agent stacks can also look at products such as Grok Voice Agent Builder. MAI-Transcribe-2-Streaming is the listening layer, not the entire agent.
Live captions and dictation
This is the clearest use case. Teams can test whether words appear quickly enough for a demo, webinar or support interaction and whether names, product terminology, accents and mixed-language speech remain readable. The useful metrics are correction effort, important-word errors and whether users can follow the transcript comfortably without waiting for it to settle.
Customer research
Sales calls, onboarding conversations, interviews and support recordings can become valuable research once transcripts are searchable. Marketing teams can aggregate them to identify recurring objections, competitor mentions, product confusion and the language customers naturally use.
For this workflow, batch transcription should usually be the first comparison point. One prospect mentioning a problem is an anecdote. The same objection appearing across dozens of accounts or deal stages is a pattern worth investigating, and that analysis rarely requires the transcript to arrive while the person is still speaking.
What marketing teams should watch next
- General availability and an SLA. That would materially change the risk calculation for customer-facing systems.
- Pricing after 2026. The current streaming and batch rates are introductory.
- Richer streaming metadata. Confidence scores, word-level timing and speaker information would make the model more useful for analytics-heavy applications.
- Performance on real company audio. Accents, interruptions, product names and noisy calls will matter more than a composite leaderboard once teams start testing it.
OneMetrik Takeaway
MAI-Transcribe-2-Streaming is worth evaluating when faster access to spoken words changes what the application can do while the conversation is still happening. The benchmark is impressive, but the benchmark is not the business case.
For live interaction, test the model with approved audio that includes accents, interruptions and company-specific terminology. Measure important-word errors, response delay, correction effort and cost per completed workflow. For recorded customer analysis, start by comparing batch transcription and check whether real-time output actually changes the downstream decision.
Pay for real-time transcription only when the workflow benefits from real time. Otherwise, a faster transcript can simply become a more expensive way to produce notes nobody acts on.