MAI-Voice-2.1-Flash

// Overview

MAI-Voice-2.1-Flash is the fast variant of MAI-Voice-2.1, announced by Microsoft AI on 1 October 2026 together with MAI-Voice-2.1 and the streaming speech-to-text model MAI-Transcribe-2-Streaming. Microsoft says it "supports the same languages, and cross-language speakers, as MAI-Voice-2.1" — 23 languages and 26 locales — "but it's been leveled up for high-volume, latency-sensitive workloads."

The launch post puts numbers on that: the model "can generate 45s of audio, with an end-to-end latency, of a mere 150ms", and Microsoft says it delivers 55% faster model inference and is about 60% cheaper than comparable models. Microsoft Learn positions it for real-time voice agents and assistants, low-latency call-centre and IVR flows, and multilingual interactive experiences.

Microsoft pitches it as the speaking half of a voice agent: "Pairing MAI-Transcribe-2-Streaming with MAI-Voice-2.1-Flash buys time back on both ends" — the streaming model transcribes the caller quickly and Flash answers at conversational speed. Like MAI-Voice-2.1, it supports fine-grained emotion and style control through SSML `mstts:express-as`, a library of licensed prebuilt voices, and gated instant voice cloning from a 5–60 second consented reference clip.

On Microsoft Foundry the model is in public preview, served globally from 14 Azure regions. It is called through the Azure Speech SDK with SSML — a prebuilt voice takes the model as a suffix, such as `en-US-Harper:MAI-Voice-2.1-Flash` — and through Azure Voice Live. Microsoft's launch post lists a price of $15 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

• Generates 45 seconds of audio with 150ms end-to-end latency, per Microsoft's launch post
• 55% faster model inference, by Microsoft's account
• $15 per 1M characters, which Microsoft says is about 60% cheaper than comparable models
• Same 23 languages, 26 locales and cross-language voices as MAI-Voice-2.1
• Emotion and style control through SSML `mstts:express-as`, plus gated instant voice cloning from a 5–60 second reference clip
• Available in Azure Voice Live as well as the Azure Speech SDK, served from 14 Azure regions

// Best for

• Reach for it for real-time voice agents and assistants, the workload Microsoft built it for
• Reach for it for low-latency call-centre and IVR flows, a use Microsoft Learn lists for the model
• Reach for it as the speech output of a voice agent paired with MAI-Transcribe-2-Streaming, as Microsoft recommends
• Reach for MAI-Voice-2.1 instead for audiobooks and long-form narration, where fidelity matters more than speed

// 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.