MAI-Voice-2.1 is Microsoft AI’s text-to-speech pair, announced on October 1, 2026 alongside MAI-Transcribe-2-Streaming. The two models are meant to be used together: Transcribe-2-Streaming handles the “hear” half of a voice agent, and MAI-Voice-2.1 or its low-latency sibling MAI-Voice-2.1-Flash handles the “speak” half.
The distinctive claim is multilingual identity. Microsoft says a single voice can speak English, then Mandarin, then German and remain unmistakably the same speaker, picking up each language’s native accent instead of dragging one accent across all of them. Both models are in public preview on Azure.
The two models
| MAI-Voice-2.1 | MAI-Voice-2.1-Flash | |
|---|---|---|
| Positioning | Highest fidelity, most expressive; long-form narration | Built for volume — fast, low-latency, lower cost |
| Model inference latency | ~550 ms | ~45 ms |
| Price | $22 per 1M characters | $15 per 1M characters |
| Languages | 23 languages, 26 locales | Same as MAI-Voice-2.1 |
| Best for | Audiobooks, content creation, voice-over, lectures | Call-center agents, voice assistants, IVR |
| Emotion control | Yes (SSML) | Yes (SSML) |
| Zero-shot voice prompting | Yes (gated) | Yes (gated) |
Microsoft’s launch post adds that Flash can generate 45 seconds of audio with a 150 ms end-to-end latency, delivers 55% faster model inference and is about 60% cheaper than comparable models. Those are vendor figures with no named comparison, so test latency on your own routes. MAI-Voice-2.1 itself is explicitly out of scope for latency-critical scenarios: it prioritizes naturalness and expressiveness.
Features
- Cross-language voice. One voice identity across all supported languages, each with a native accent, and cross-language cloning.
- Instant voice cloning (gated). Match a consented reference voice from a 5–60 second clip without fine-tuning. Built-in guardrails allow only authorized, consented voices.
- Emotion and style control with SSML
mstts:express-asand astyleattribute (joy, excitement, empathy and others). - Voice library. Licensed, curated prebuilt voices in every supported language.
- Long-form consistency. MAI-Voice-2.1 targets stable persona quality and speaker consistency across audiobooks, podcasts and lectures.
Languages
The model card states 23 languages and 26 locales. The languages named include English (US, Australia, UK, India), Italian, French, German, Hindi, Spanish (Spain, Mexico), Portuguese (Brazil, Portugal), Korean, Simplified Chinese, Turkish, Russian, Thai, Dutch, Romanian, Hungarian, Czech, Danish, Finnish, Indonesian, Polish, Swedish, Norwegian (Bokmål) and Vietnamese. Microsoft’s model page lists 28 labels under its “23 Languages” heading, so treat the exact locale count as unreconciled and test the specific locale you need.
Pricing and status
- Price: $22 per 1M characters (MAI-Voice-2.1) and $15 per 1M characters (Flash), per Microsoft AI’s model page. Azure’s speech pricing page is the authoritative billing reference.
- Rough per-minute cost: assuming 900–1,000 characters per minute of speech (our estimate), about $0.02 per minute for MAI-Voice-2.1 and $0.014 for Flash — in the same range as Gemini 3.8 Flash TTS’s launch pricing.
- Status: public preview. Microsoft Learn says the preview comes without a service-level agreement and is not recommended for production workloads.
- Usage rights: Microsoft says the models are available to third-party developers, and that Microsoft holds full licensing rights for commercial use.
Availability
- Microsoft Foundry and Azure Speech: SDK (.NET, Python, Java, JavaScript, C++) and REST. Use SSML voice names such as
en-US-Harper:MAI-Voice-2.1anden-US-Harper:MAI-Voice-2.1-Flash. - Azure Voice Live, the MAI Playground (including a live-agent demo called Chatter), OpenRouter and Vercel; LiveKit is listed as coming soon. Microsoft also lists Copilot Audio Expressions.
- Regions: accessible globally, served from 14 Azure regions, including France Central, East US, West Europe, Sweden Central and Japan East.
How it compares
- vs Gemini 3.8 Flash TTS: Google offers 130 languages, prompt-based voice design, two-speaker scene staging and a free tier; MAI-Voice-2.1 offers one voice across 23 languages with native accents and a tight pairing with Microsoft’s streaming transcription.
- vs Eleven v4: ElevenLabs leads independent blind votes with 90+ languages and a mature ecosystem, at $40–$80 per 1M characters; MAI-Voice is cheaper on paper ($15–$22) but unmeasured and still in preview.
- vs MiniMax Speech: MiniMax clones a voice from about ten seconds of audio at $1.50 per voice, but charges $60–$100 per 1M characters; MAI-Voice is cheaper per character, builds consent guardrails into cloning and is delivered through Azure’s Speech SDK and Foundry.
- vs Qwen-Audio 3.1: Alibaba’s lineup spans ASR, TTS and realtime in one family; Microsoft splits transcription and voice into separate MAI models.
Limitations
- Preview status. No SLA; capabilities and prices may change.
- No published benchmark numbers for the voice models, only qualitative claims. Neither is on Artificial Analysis’s Speech Arena yet; its predecessor MAI-Voice-1 ranks #72 (Elo 1033), against 1330 for the leader.
- Gated cloning requires consent verification and may not be enabled for every account.
- Latency trade-off. The flagship model is not meant for real-time agents; use Flash.
See the audio models directory for other speech models.
Best for
- Voice agents, call-center and IVR flows (Flash), paired with MAI-Transcribe-2-Streaming
- Audiobooks, podcasts, lectures and voice-over (MAI-Voice-2.1)
- Multilingual assistants that answer in the user's language with one consistent voice
- Brand-voice cloning across markets, with consent verification
Pros & cons
Strengths
- One voice identity across all 23 languages (26 locales) with a native accent in each, so a brand voice or tutor can switch languages mid-conversation
- Two tiers — MAI-Voice-2.1 for fidelity and long-form consistency (audiobooks, voice-over), Flash for ~45 ms model inference and 150 ms end-to-end latency in agents
- Flash is priced at $15 per 1M characters; Microsoft says it is about 60% cheaper than comparable models with 55% faster inference
- Instant voice cloning from a 5–60 second clip with no fine-tuning, behind consent guardrails, plus SSML emotion control (joy, excitement, empathy and more)
Limitations
- Public preview — no SLA and not recommended for production workloads by Microsoft's own docs
- Microsoft publishes no independent benchmark numbers for the voice models (it claims "leading expressiveness and naturalness"), unlike its Transcribe-2 launch
- Voice cloning is gated, and the main model prioritizes naturalness over latency (~550 ms inference)
- Text-only input via SSML; the model page lists 28 language labels under a "23 languages" headline, while the model card states 23 languages and 26 locales
Frequently asked questions
What is MAI-Voice-2.1?
MAI-Voice-2.1 is Microsoft AI's in-house prompted text-to-speech model, announced October 1, 2026 together with MAI-Voice-2.1-Flash and MAI-Transcribe-2-Streaming. It generates expressive speech in 23 languages and 26 locales and can keep one voice identity across every supported language.
How much does MAI-Voice-2.1 cost?
Microsoft AI lists $22 per 1M characters for MAI-Voice-2.1 and $15 per 1M characters for MAI-Voice-2.1-Flash. At around 900–1,000 characters per minute of speech, that is roughly $0.02 and $0.014 per minute. Check Azure's speech pricing page for billing, since the models are in public preview.
What is the difference between MAI-Voice-2.1 and MAI-Voice-2.1-Flash?
MAI-Voice-2.1 is the highest-fidelity model, built for long-form narration, audiobooks and voice-over; the model page lists about 550 ms inference latency. Flash targets high-volume, latency-sensitive work such as call-center agents, assistants and IVR, with about 45 ms inference latency and Microsoft's claim of 150 ms end-to-end for generating 45 seconds of audio. Both share the same 23 languages.
Can MAI-Voice-2.1 clone my voice?
Yes, in a gated way. It matches a consented reference voice from a 5–60 second clip without fine-tuning, including across languages, with built-in guardrails that only allow authorized, consented voices.
Where can I use MAI-Voice-2.1?
Through Microsoft Foundry and Azure Speech (SDK and REST, using SSML voice names such as en-US-Harper:MAI-Voice-2.1), Azure Voice Live, the MAI Playground, OpenRouter and Vercel, with LiveKit listed as coming soon. Azure serves it globally from 14 regions.
Is MAI-Voice-2.1 production-ready?
Not by Microsoft's own wording. Its Learn page describes the feature as public preview, without a service-level agreement and not recommended for production workloads. Plan for the status to change before you commit to it.
Sources
- Microsoft AI — MAI-Voice-2.1 model page
- Microsoft AI — Our first streaming transcription model debuts at no. 1 on Artificial Analysis (launch post, Oct 1, 2026)
- Microsoft Learn — MAI-Voice-2.1 and MAI-Voice-2.1-Flash (Speech Service)
- Microsoft AI — MAI-Voice-2.1 model card
- Microsoft AI — MAI-Voice-2.1-Flash model card