Skip to content
← Models
ClosedMicrosoft/10 trustMetrics-checked
MAI-Voice-2.1 logo

MAI-Voice-2.1

Microsoft's most expressive text-to-speech model — 23 languages with native-accent cross-language voice identities.

Visit official page →

About

MAI-Voice-2.1, released October 1, 2026 alongside MAI-Transcribe-2-Streaming and the fast MAI-Voice-2.1-Flash, is Microsoft AI's multilingual text-to-speech model for voice agents. Microsoft describes it as its most expressive such model yet: natural, expressive speech in 23 languages, with cross-language voice identities that keep native accents when switching languages.

Both voice models can replicate voices from a few seconds of reference audio, gated behind access approval and consent requirements. The Flash variant is built for responsive, high-volume use cases at $15 per million characters versus $22 for the full model. All three launched in public preview through Microsoft Foundry and Azure AI Speech, and Vercel's AI Gateway carries the MAI voice and transcription lineup.

No independent quality benchmarks have been published, so the expressiveness claims are Microsoft's own until third parties evaluate the voices.

Specs

Context window

N/A — speech synthesis model

Pricing

$22 per 1M characters (Voice-2.1-Flash: $15 per 1M)

Free quota

none published

API access

Microsoft Foundry / Azure AI Speech (public preview); Vercel AI Gateway

Weights

Closed — API only

Released

October 1, 2026

Benchmarks

No independent quality benchmarks found; provider describes it as its most expressive TTS model.

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

voice-agentsmultilingual-ttscustomer-servicecontent-narration

Trust score — 7/10

/10 trustMetrics-checked

Launch facts (October 1, 2026; 23 languages; $22/$15 per 1M characters; Foundry public preview) corroborated across independent outlets. Docked for: (1) no independent quality benchmarks — expressiveness claims are Microsoft's own; (2) public preview with no SLA; (3) voice cloning requires access approval and consent, which limits availability.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
7.0/10
Maintenance · recency, releases, issue load
8.0/10
Openness · license + docs
3.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · Adoption 7: launched Oct 1, 2026 with Transcribe-2-Streaming; multiple outlet corroboration of specs and pricing
  • · Maintenance 8: Foundry public preview; Flash variant for high-volume tier
  • · Openness 3: proprietary API-only, no weights
  • · Manual 7: 23 languages and $22/$15 per 1M chars verified; expressiveness claims are Microsoft's own, no independent benchmarks

More in the mai-voice family

Sources