Voice is becoming the interface, and Microsoft just shipped two models built to power it. As part of the seven-model MAI family announced at Build 2026, MAI-Voice-2 handles expressive text-to-speech across 15 languages, and MAI-Transcribe-1.5 handles speech-to-text across 43 languages. Together they form a complete audio loop: listen, understand, and respond in a natural voice.

The headline numbers are striking. MAI-Transcribe-1.5 can transcribe an hour of audio in under 15 seconds at a 2.4% Word Error Rate, and MAI-Voice-2 produces speech that listeners preferred over its predecessor 72% of the time, with voice cloning from as little as five seconds of reference audio. For anyone building call centers, audiobooks, accessibility tools, or voice agents, this is a serious production stack.

This guide covers both models in depth: capabilities, benchmarks, keyword biasing, consent guardrails, and how to build voice agents on Microsoft Foundry. For the full MAI lineup, see our Microsoft MAI models developer guide.

1MAI-Voice-2: Expressive Text-to-Speech

MAI-Voice-2 is Microsoft's most expressive text-to-speech model to date, and a significant jump from MAI-Voice-1 across the dimensions that matter for production voice: fidelity, language coverage, speaker consistency, and emotional range. It expands from English-only to 15 languages while maintaining the same naturalness and expressiveness as English.

The standout features for developers:

• Granular emotion control via emotion tags such as sad, whispered, or excited, plus role styles like motivational trainer or sports commentator
• Stable speaker identity across long-form content, so a single voice holds up across an entire audiobook, podcast, or lecture
• Code-switching for select language pairs such as Hindi-English and Spanish-English, switching mid-sentence without losing prosody or speaker identity
• Strong preference results - MAI-Voice-2 was preferred over MAI-Voice-1 72.1% of the time across roughly 2,500 listening tests

2Voice Cloning & Consent Guardrails

MAI-Voice-2 supports zero-shot voice prompting: with just 5 to 60 seconds of reference audio, developers can create a custom voice in Microsoft Foundry across all supported languages, with no retraining or fine-tuning required. That makes it practical for companies to bring a consistent brand voice into their products without maintaining a separate voice model.

Microsoft pairs this capability with strict consent controls. Consent is enforced at the system level, so only authorized, licensed voices can be synthesized in production, and Microsoft states no unlicensed voice cloning is possible. Access to the voice-creation feature requires an application.

3MAI-Transcribe-1.5: Speech-to-Text

MAI-Transcribe-1.5 is Microsoft's most accurate multilingual speech-to-text model, with best-in-class Word Error Rate across 43 languages on the FLEURS benchmark. It expanded coverage from 25 to 43 languages without sacrificing accuracy, and Microsoft positions it as the fastest, most efficient, and most cost-effective transcription model among the hyperscalers.

The speed jump is the headline: transcribing an hour of audio in under 15 seconds is up to five times faster on long audio than models like Gemini 3.1, Scribe v2, and GPT-4o-Transcribe. It is also optimized for messy real-world conditions such as noisy backgrounds, which is exactly where many transcription models fall down.

4Keyword Biasing for Domain Vocabulary

The feature most likely to matter in production is keyword biasing. Generic transcription models routinely mangle the words that matter most: people and product names, medical terms, internal acronyms, and customer-specific vocabulary. MAI-Transcribe-1.5 lets you supply a list of domain-specific keywords and biases its predictions toward them.

Crucially, the model does not blindly force matches. It uses the surrounding context to decide when biasing should apply, so it improves recognition of specialized vocabulary while keeping accuracy on general speech. Microsoft reports a 30% reduction in Word Error Rate on the FLEURS multilingual benchmark when keyword biasing is used. In its example, names like Aoife, Xochitl, and Soren that a baseline model garbled were transcribed correctly once supplied as keywords.

# Transcription with keyword biasing (illustrative)
POST https://<your-foundry-endpoint>/transcribe
Authorization: Bearer <FOUNDRY_API_KEY>

{
  "model": "mai-transcribe-1.5",
  "audio_url": "https://.../call-recording.wav",
  "language": "en",
  "keywords": [
    "Aoife", "Xochitl", "Soren", "Niamh",
    "MAI-Transcribe", "Foundry", "CPT-99457"
  ]
}

Microsoft also flagged what is coming next: speaker diarization (who said what in multi-speaker audio), a native streaming API for real-time transcription, and continued language expansion. The streaming API in particular will matter for live voice agents, since the current model is batch-first.

5Building a Voice Agent Loop

The two models are designed to compose. A typical voice agent loop chains transcription, reasoning, and speech synthesis, with a language model in the middle that you can swap as needed. Microsoft's own DuoAI demo shows this exact pattern, combining MAI-Transcribe-1.5, MAI-Voice-2, and MAI-Image-2.5 into a multi-agent conversation.

Until the native streaming transcription API ships, design real-time agents around the batch-first model with short audio chunks, and budget for the latency that introduces. For the orchestration and guardrail side of agent design, see our multi-agent orchestration patterns guide.

6Access & What's Next

Both models are available in Microsoft Foundry and the MAI Playground. MAI-Voice-2 is now in Azure Foundry and integrating into VS Code and the Dynamics 365 Contact Center. MAI-Transcribe-1.5 is in Foundry and being integrated into Copilot, Teams, GitHub, and Dynamics 365 Contact Center. A MAI-Voice-2-Flash variant is coming for lower-cost, ultra-efficient synthesis.

• MAI-Voice-2 model card and Foundry docs for the TTS API and cookbook
• MAI-Transcribe-1.5 model card and Foundry docs for the transcription API, including keyword biasing
• DuoAI on MAI Playground to try both models in a live multi-agent conversation

7Why Lushbinary for Voice AI

Production voice systems are deceptively hard. Latency budgets, consent management, multilingual edge cases, and graceful failure handling all have to work together. Lushbinary builds voice and speech applications end-to-end, from contact-center automation to accessibility tools and branded assistants.

• Voice agent development - full transcribe, reason, and speak loops built on MAI or a routed mix of models
• Consent & compliance - voice-cloning consent workflows and audit trails that hold up to scrutiny
• Multilingual deployment - tuning for accuracy and naturalness across the languages your users actually speak
• Foundry & Azure integration - secure, scalable deployment with monitoring and cost controls

8Frequently Asked Questions

📚 Sources

• Microsoft AI - Introducing MAI-Voice-2
• Microsoft AI - Introducing MAI-Transcribe-1.5
• Microsoft AI - Building a hill-climbing machine

Content was rephrased for compliance with licensing restrictions. Language coverage, accuracy, speed, and preference figures sourced from official Microsoft AI announcements as of June 2, 2026. All figures are vendor-reported and may change - always verify on Microsoft's website.