MAI Playground, Microsoft AI’s public test bench for its in-house models, gained three new audio models on 1 October 2026. MAI-Transcribe-2-Streaming turns live speech into text as people talk, and two new voice models, MAI-Voice-2.1 and the faster MAI-Voice-2.1-Flash, turn text back into speech in 23 languages. Microsoft also added a demo called Chatter, a voice assistant built from the new models, so anyone can hear the pieces working together.
TestingCatalog spotted the additions the same day, quoting Microsoft’s own summary: “MAI-Transcribe and MAI-Voice models: accurate, fast, low cost, and chart-topping audio understanding and generation for building the best conversational voice agents.” The models are also available to developers through Microsoft Foundry and partners.
This article explains what is new in MAI Playground, what each model does and costs, how the streaming transcription model ranks against rivals, how quickly Microsoft’s voice line has moved this year, and what developers in the UK should check before building on it. For the batch model that came first, see our earlier report on MAI-Transcribe-2.
Table of contents
- What Microsoft Added to MAI Playground
- MAI-Transcribe-2-Streaming: Real-Time Speech to Text
- MAI-Voice-2.1 and MAI-Voice-2.1-Flash
- How Fast Microsoft’s Voice Line Has Moved
- Why Microsoft Is Building Its Own Voice Models
- Where to Use the Models Beyond MAI Playground
- How to Test the New Models in MAI Playground
- How the New Models Compare
- What It Means for UK Developers
- MAI Playground FAQ
- References and Further Reading
What Microsoft Added to MAI Playground
Microsoft announced the three models in a post titled “Our first streaming transcription model debuts at no. 1 on Artificial Analysis”, published at 16:00 UTC on 1 October.
Three models in one release
MAI-Transcribe-2-Streaming is Microsoft’s first real-time speech recognition model. MAI-Voice-2.1 is its most capable text-to-speech model, and MAI-Voice-2.1-Flash is the low-latency version built for volume. Microsoft frames them as the two ends of a voice agent: one model listens, the other speaks.
Chatter, the new demo
Chatter is a new voice assistant demo inside MAI Playground, labelled Beta and “Powered by voice & transcribe”. Microsoft says it built Chatter “to show these models working together in a live agent.” It replaces the idea of testing each model separately with a single spoken conversation.
What the MAI Playground menu shows now
When we checked on 2 October, the MAI Playground model picker listed MAI-Transcribe-2, MAI-Voice-2.1, MAI-Image-2.6 and MAI-Thinking-1, with Chatter under AI Experiences. Microsoft’s own model pages for MAI-Transcribe-2 and MAI-Voice-2.1 were updated on 1 October with “Try in playground” buttons.
| Model | Job | Price | Languages |
|---|---|---|---|
| MAI-Transcribe-2-Streaming | Real-time speech to text | $0.54 per audio hour (introductory, to end of 2026) | 60 |
| MAI-Voice-2.1 | Expressive text to speech | $22 per 1M characters | 23 |
| MAI-Voice-2.1-Flash | Low-latency text to speech | $15 per 1M characters | 23 |
| Chatter (demo) | Voice assistant using the models above | Free in MAI Playground | Not stated |
MAI-Transcribe-2-Streaming: Real-Time Speech to Text
The streaming model is the headline. Batch transcription waits for a whole file; streaming transcription has to produce words while someone is still talking, which is much harder to do accurately.
First words in about 100 milliseconds
Microsoft says the model produces its first guesses, known as partials, “in just over 100ms of receiving audio”, then revises them as more context arrives and commits a stable transcript. That lets a voice agent start reasoning or calling tools before the speaker finishes. Microsoft’s internal tests show words appearing “2x faster” than its closest competitor for dictation and subtitles.
Sixty languages with automatic detection
The streaming model covers 60 languages, matching the batch MAI-Transcribe-2, and supports automatic, continuous language detection. That matters for call centres and meetings where people switch languages mid-conversation.
How it ranks on Artificial Analysis
Microsoft cites the Artificial Analysis streaming speech-to-text leaderboard as of 28 September 2026. There, MAI-Transcribe-2-Streaming had the lowest final word error rate of the streaming models shown, 2.50%, and a first-partial error rate of 2.80%. It returned a final transcript 0.130 seconds after speech ended.
Final word error rate on Artificial Analysis streaming leaderboard, 28 Sep 2026, lower is better (bar length = WER ÷ 5.25%)
Microsoft beats its own older service
One detail stands out in Microsoft’s chart: its existing Azure real-time speech service sits at 5.25%, more than double the new model’s 2.50%. Teams already using Azure Speech for live captions have a clear reason to test the new model.
Speed is not uniform
Low error is only half of a streaming result. On the same chart, some rivals return a final transcript faster: Soniox v5 Real-Time at 0.054 seconds and Cartesia Ink-2 at 0.067 seconds, against Microsoft’s 0.130. Microsoft’s claim is that it sits on the best balance of accuracy and speed, not that it is the fastest.
Price compared with the batch model
The streaming model costs an introductory $0.54 per hour of audio until the end of the year. The batch MAI-Transcribe-2 launched on 3 September at $0.10 per hour. Real-time transcription therefore costs 5.4 times as much per hour ($0.54 ÷ $0.10), which is normal for streaming but worth budgeting for if you only need recordings transcribed after the fact.
MAI-Voice-2.1 and MAI-Voice-2.1-Flash
The two voice models share the same languages and voices but are tuned for different jobs, and both now appear in MAI Playground.
Twenty-three languages, one voice
Microsoft says MAI-Voice-2.1 supports 23 languages and lets “one single voice” speak all of them “with a truly native accent”, so a brand can keep the same voice everywhere. The model page lists English, Italian, French, German, Hindi, Spanish, Portuguese, Korean, Chinese (Simplified), Turkish, Russian, Thai, Dutch, Romanian, Hungarian, Czech, Danish, Finnish, Indonesian, Polish, Swedish, Norwegian and Vietnamese. Japanese and Arabic are not on the list.
Counting the locales
The blog says 23 languages and 26 locales. The model page’s list has 28 entries, because English appears in four variants (US, Australia, UK and India) and Spanish and Portuguese in two each. Microsoft has not explained the difference, so check the exact locale you need in MAI Playground before you commit.
Prices have not moved
MAI-Voice-2.1 costs $22 per million characters and Flash costs $15. When Microsoft launched MAI-Voice-2-Flash in July, it said Flash was 32% cheaper than MAI-Voice-2 at $15, which implies MAI-Voice-2 was about $22 ($15 ÷ 0.68). So the 2.1 upgrade adds eight languages at the same price.
Latency figures
Microsoft’s model page lists model inference latency of about 550 milliseconds for MAI-Voice-2.1 and about 45 milliseconds for Flash. The blog gives Flash “an end-to-end latency” of 150 milliseconds and says it delivers “55% faster model inference” and is “~60% cheaper than comparable models.” Inference time and end-to-end time measure different things, so expect real calls to land nearer the larger figure.
| Feature | MAI-Voice-2.1 | MAI-Voice-2.1-Flash |
|---|---|---|
| Model inference latency | About 550 ms | About 45 ms |
| Price per 1M characters | $22 | $15 |
| Languages | 23 | 23 |
| Emotion control and voice prompting | Yes | Yes |
| Best for (Microsoft) | Audiobooks, content, voice-over | Call centres, assistants, IVR |
What listeners thought
Microsoft reports two listener tests. In a blind side-by-side comparison of 5,032 judgements, MAI-Voice-2.1 was preferred 58.9% of the time against 41.1% for MAI-Voice-2. In a test with 4,000 listeners, 50.3% rated the MAI voices as equally or more human-like than human recordings. Both are Microsoft’s own studies, and the second combines results for both 2.1 models.
Cloning with consent guardrails
Both voice models can clone a voice “using just a few seconds of reference audio” in every supported language, and Microsoft says they have “built-in consent guardrails that prevent misuse.” For the legal side of voice cloning, see our report on a Japanese court ruling on voice rights.
How Fast Microsoft's Voice Line Has Moved
The additions to MAI Playground cap a year in which Microsoft has shipped a new audio model roughly every month. Microsoft AI, led by Mustafa Suleyman, has steadily reduced its reliance on partner models for speech.
| Date | Release | Key fact |
|---|---|---|
| 2 April 2026 | MAI-Transcribe-1 | 25 languages |
| 4 June 2026 | MAI-Voice-2 | English-only to 15 languages |
| 24 July 2026 | MAI-Voice-2-Flash public preview | $15 per 1M characters, 2x faster |
| 3 September 2026 | MAI-Transcribe-2 | 60 languages, $0.10 per hour |
| 1 October 2026 | MAI-Transcribe-2-Streaming, MAI-Voice-2.1, MAI-Voice-2.1-Flash | Real-time, 23 voice languages, added to MAI Playground |
Languages supported, by release (bar length = languages ÷ 60)
On these figures, voice coverage grew by about half in four months (from 15 to 23 languages), while transcription coverage more than doubled in five (from 25 to 60).
Why the pace matters
Each release has gone into Microsoft’s own products before or alongside developers. In July Microsoft said a MAI transcription model was replacing the previous model in Dragon Copilot, its clinical documentation tool used by 170,000 medical providers. Our report on Microsoft’s draft code of conduct covers the principles it says govern these releases.
Why Microsoft Is Building Its Own Voice Models
The steady stream of releases into MAI Playground is part of a bigger strategy. Microsoft has long relied on partner models, above all OpenAI’s, inside Copilot. Its MAI line gives it in-house alternatives that it controls, prices and ships on its own schedule.
Products first, developers second
Each MAI audio model has gone into Microsoft’s own products as well as to developers. Microsoft said earlier MAI-Transcribe versions were used in Copilot Voice and Teams, and that MAI-Voice-2 was being built into VS Code and the Dynamics 365 Contact Center. When a model performs in those products, it then appears in MAI Playground and Foundry for everyone else.
Price as a weapon
Microsoft is competing hard on price. MAI-Transcribe-2 launched at $0.10 per audio hour, and the streaming model at an introductory $0.54. Both introductory prices run only until the end of 2026, so budget for a possible increase next year rather than assuming today’s rate will last.
What it means for Azure Speech customers
Azure’s existing real-time speech service scored 5.25% on the same streaming leaderboard, against 2.50% for the new model. Microsoft has not said whether the older service will be retired, but customers using it for captions or call transcription now have a clearly better option from the same vendor. Testing it in MAI Playground first costs nothing.
A small lab moving quickly
Microsoft describes its MAI team as “a lean, fast-moving lab” with compute that “is ramping quickly and extensively.” Five audio releases in six months suggest that pace will continue, so treat any comparison you run in MAI Playground today as a snapshot rather than a final verdict.
Where to Use the Models Beyond MAI Playground
MAI Playground is for listening and testing. For production, Microsoft lists several routes.
Microsoft Foundry and Azure Voice Live
All three models are available through Microsoft Foundry, and Microsoft points voice-agent builders to Azure Voice Live, its speech-to-speech service. The MAI-Voice-2.1 page also links to Azure Speech and to Copilot Audio Expressions.
OpenRouter, Vercel and LiveKit
MAI-Voice-2.1 and Flash are available through OpenRouter. Microsoft also lists Vercel, and says LiveKit support is coming soon. That makes it easier to swap MAI voices into an existing agent stack without moving everything to Azure.
Building a voice agent loop
Microsoft’s own advice is to pair MAI-Transcribe-2-Streaming with MAI-Voice-2.1-Flash. “A voice agent is a loop. It has to hear, understand, decide, and speak,” its post says, and every component “either buys you time in that window… or spends it.” Saving time on listening and speaking leaves more time for the agent to reason and use tools.
Use cases Microsoft names
Microsoft lists customer service agents that act before the caller finishes, multilingual assistants that reply in the language they are addressed in, and interactive learning with distinct speakers for tutoring and role-play. Each depends on natural language processing between the two audio models, typically a large model that decides what to say.
How to Test the New Models in MAI Playground
MAI Playground is free to use, so the quickest way to judge the new models is to try them on your own material. Here is a simple plan.
Start with Chatter
Open Chatter in MAI Playground and hold a normal conversation. Notice how quickly it replies, whether it waits for you to finish, and what happens when you interrupt it mid-sentence. That tells you more about the streaming transcription and Flash voice working together than any benchmark chart.
Test transcription on your own audio
Record two or three minutes that look like your real use: a phone call, a meeting, a support conversation. Include regional accents, technical terms and some background noise. Run it through MAI Playground and count the errors by hand. If your current provider can process the same clip, compare the two results line by line.
Compare the two voices
Give MAI-Voice-2.1 and MAI-Voice-2.1-Flash the same script and listen to both. Try a sentence with a product name, a long number and a question, then ask the same voice to repeat it in a second language. The difference in pacing and expression will show whether you need the full model or whether Flash is good enough.
Know what MAI Playground cannot tell you
MAI Playground shows quality, not operations. It will not tell you about service levels, data retention, regional hosting or costs at volume. Those live in Microsoft Foundry and Azure documentation and in your contract, so treat a good MAI Playground session as the start of an evaluation, not the end of one.
How the New Models Compare
MAI Playground makes it easy to hear Microsoft’s models, but buyers will want to compare them with rivals they may already use.
Streaming transcription rivals
On the Artificial Analysis chart Microsoft published, the nearest rivals on accuracy were xAI’s Grok Voice Transcribe 2.0 and Meta’s Muse Voice Transcribe, followed by Cartesia, ElevenLabs Scribe v2 Realtime, OpenAI’s GPT Live Transcribe and Google’s Gemini 3.5 Transcribe Live. The gap between first place and ElevenLabs in seventh is 1.09 percentage points (3.59% minus 2.50%), so test on your own audio. Our report on Gemini 3.5 Transcribe covers Google’s batch model.
Voice rivals
ElevenLabs launched Eleven v4 and v4 Turbo on 28 September and topped the TTS Arena leaderboard, as we reported in our story on ElevenLabs Eleven v4. Microsoft’s pitch against that is price and language breadth per voice, rather than a leaderboard win.
| Streaming model | Final WER | First partial WER | Time to final |
|---|---|---|---|
| MAI-Transcribe-2-Streaming | 2.50% | 2.80% | 0.130 s |
| Grok Voice Transcribe 2.0 | 2.73% | 3.36% | 0.490 s |
| Muse Voice Transcribe | 3.06% | 3.57% | 0.163 s |
| ElevenLabs Scribe v2 Realtime | 3.59% | 3.59% | 0.141 s |
| Gemini 3.5 Transcribe Live | 4.00% | 5.77% | 0.395 s |
Figures are from the Artificial Analysis data Microsoft embedded in its 1 October post, dated 28 September 2026.
What It Means for UK Developers
The new models are easy to try in MAI Playground, but production use raises the usual questions.
Voice data is personal data
Recordings of people’s voices are personal data under UK GDPR, and a voice used to identify someone can be special category biometric data. If you transcribe calls, tell callers, set retention periods and check where processing happens in your Azure or Foundry region.
Cloning needs clear consent
Cloning a staff member’s or actor’s voice needs written consent that covers how long and where the voice will be used. Microsoft’s guardrails help, but the responsibility stays with you. For the security side of call-centre voice, our IT security team can help you plan against cloned-voice fraud.
Test with your own accents
Benchmarks use standard test sets. Before switching a UK contact centre, run your own recordings through MAI Playground, including regional accents and background noise, and compare the error rate with your current provider. Our AI and machine learning team can help design that comparison.
MAI Playground FAQ
What is MAI Playground?
MAI Playground is Microsoft AI’s free web space for trying Microsoft’s in-house models, including transcription, voice, image and reasoning models, plus demos such as Chatter.
What was added to MAI Playground on 1 October 2026?
MAI-Transcribe-2-Streaming, MAI-Voice-2.1 and MAI-Voice-2.1-Flash, together with Chatter, a voice assistant demo built on the new models.
How much do the new models cost?
MAI-Transcribe-2-Streaming costs an introductory $0.54 per audio hour until the end of 2026. MAI-Voice-2.1 costs $22 and MAI-Voice-2.1-Flash $15 per million characters.
How many languages do they support?
The streaming transcription model supports 60 languages. Both voice models support 23 languages, with one voice able to speak all of them.
Is MAI-Transcribe-2-Streaming the most accurate?
On Artificial Analysis’s streaming leaderboard dated 28 September 2026, as published by Microsoft, it had the lowest final word error rate of the models shown, at 2.50%.
Where can developers use the models?
Through Microsoft Foundry, Azure Speech and Voice Live, OpenRouter (voice models), Vercel and, soon, LiveKit, as well as for testing in MAI Playground.
References and Further Reading
Our first streaming transcription model debuts at no. 1 on Artificial Analysis (Microsoft AI)
MAI-Voice-2.1 model page (Microsoft AI)
MAI-Transcribe-2 model page (Microsoft AI)
TestingCatalog on the new MAI-Transcribe and MAI-Voice releases in MAI Playground (Threads)
Introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash (Microsoft AI)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
