MAI-Voice-2.1

// Overview

MAI-Voice-2.1 is the text-to-speech model Microsoft AI announced on 1 October 2026, alongside a faster sibling, MAI-Voice-2.1-Flash, and the streaming speech-to-text model MAI-Transcribe-2-Streaming. Microsoft calls it "our strongest multilingual text-to-speech model yet", and says it has been expanded to support 23 languages and 26 locales "while enabling one single 'voice' to use all languages with a truly native accent."

Microsoft Learn describes it as the highest-fidelity, most expressive model in the MAI-Voice family, built for experiences where maximum voice quality matters, such as long-form narration and brand-defining audio. It is optimised for long-form generation with stable persona quality and speaker consistency across audiobooks, podcasts and lectures. The documentation is explicit about the trade-off: the model prioritises naturalness and expressivity over latency-critical scenarios, which is the job of MAI-Voice-2.1-Flash.

Speaking style is controlled through SSML: the `mstts:express-as` element with a `style` attribute steers emotions such as joy, excitement and empathy, and the prebuilt voices in the managed voice library each list their supported styles (for example `audiobook`, `narrator`, `whispering` or `sad`). Both MAI-Voice-2.1 models also support instant voice cloning from a short reference clip — Microsoft Learn recommends 5–60 seconds of audio — across all supported languages, but the feature is gated: only authorised, consented voices can be synthesised, and access requires approval through Microsoft's Limited Access Review.

On Microsoft Foundry the model is in public preview, served globally from 14 Azure regions through the same Azure Speech API and SDK as other Azure neural voices: a prebuilt voice name takes the model as a suffix, such as `en-US-Harper:MAI-Voice-2.1`. Microsoft's launch post lists a price of $22 per 1M characters and access through Microsoft Foundry, MAI Playground, OpenRouter, Vercel and Azure Voice Live, with LiveKit support marked as coming soon.

// Pricing

Price listed in Microsoft AI's launch post of 1 October 2026.

Pricing source ↗

// Strengths

• 23 languages and 26 locales, with one voice able to speak every supported language with a native accent, per Microsoft
• Highest-fidelity, most expressive model in the MAI-Voice family, according to Microsoft Learn
• Optimised for long-form generation with stable persona quality and speaker consistency across extended content
• Fine-grained emotion and style control through SSML `mstts:express-as`
• Gated instant voice cloning from a 5–60 second reference clip, with built-in consent guardrails
• Uses the standard Azure Speech REST API and SDKs (Python, C#, JavaScript, Java), served from 14 Azure regions

// Best for

• Reach for it for audiobooks and podcasts, where Microsoft Learn says it keeps a speaker consistent across long-form content
• Reach for it for educational content and lectures that need expressive narration in several of its 23 languages
• Reach for it for voice-overs and brand-defining audio, including a consented clone of a brand's own voice
• Reach for MAI-Voice-2.1-Flash instead when latency matters more than maximum fidelity, such as live voice agents

// How to access

// MAI-Voice (text-to-speech) — every version

The full lineage of the MAI-Voice (text-to-speech) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.