Microsoft Releases MAI-Voice-2 Text-to-Speech Model in Free Public Playground

Microsoft has launched its new in-house text-to-speech model, MAI-Voice-2, in a free public web-based sandbox called the MAI Playground. While the model represents a major leap in multilingual synthesis and zero-shot voice cloning, access to the experimental sandbox is currently geofenced and restricted to users located in the United States. To prevent the proliferation of unauthorized deepfakes, Microsoft is routing production-level voice deployment through a strict system-level consent and Azure review process.

Multilingual Capabilities and Expressive Speech Controls

Text-to-speech (TTS) technology has transitioned from traditional speech synthesis, which stitched pre-recorded syllables together into robotic output, to modern neural networks. Modern neural models analyze written text globally, predicting the pitch, cadence, and emotion that a human speaker would naturally use.

MAI-Voice-2, developed by Microsoft AI, expands upon its predecessor, MAI-Voice-1, which was restricted solely to English. The updated model natively supports 15 languages, enabling it to navigate diverse linguistic structures including tonal, pitch-accent, and stress-timed speech patterns. Beyond basic translation, the model supports “code-switching,” allowing it to seamlessly transition between two languages—such as Spanish and English or Hindi and English—within a single sentence without losing the speaker’s vocal identity.

A defining feature of MAI-Voice-2 is its expressive inline control. Rather than relying on external knobs or complex dial interfaces, developers and users can insert inline emotion tags directly into the text. The synthesis engine adjusts its style and tone at the exact moment the tag appears. Supported styles and emotions include:

• Emotion tags: [whispered], [excited], [sad], and [embarrassed]
• Role styles: Motivational trainer and sports commentator

For example, a synthesized dialogue can be prompted like this:

[whispered] I shouldn't tell you, but the release is already finished.
[excited] It really works, the synthesized audio is indistinguishable from my real voice!
[embarrassed] Sorry, I think I got the pronunciation wrong again.

Zero-Shot Cloning and Reference Audio Requirements

Traditional custom voice creation required hours of high-quality recording session data and computationally intensive model training. MAI-Voice-2 uses zero-shot voice cloning (also referred to as voice prompting) to instantly replicate a human voice. This method maps the voice’s acoustic characteristics directly onto the TTS synthesis pipeline without requiring any localized model fine-tuning or retraining.

To clone a voice, the model requires a brief audio clip of the target speaker. Microsoft’s recommendations for reference audio include:

1. Length: A sample of 5 to 60 seconds (though the Azure platform allows up to 120 seconds for verification).
2. Quality: Clean, background-noise-free recordings to accurately capture vocal timbre and rhythm.
3. Application: Once analyzed, the system maps these unique properties across all 15 supported languages, allowing the cloned voice to speak languages the original speaker may not know.

Subjective Human Preference and Indistinguishability Benchmarks

Evaluating neural speech engines requires listening trials because algorithmic metrics like Word Error Rate (WER) cannot fully capture warmth, rhythm, or conversational flow. Microsoft conducted extensive internal listening tests to benchmark MAI-Voice-2 against older models and real human speakers.

• Comparison to Predecessor: In side-by-side preference tests across approximately 2,500 trials, evaluators preferred speech synthesized by MAI-Voice-2 over MAI-Voice-1 72% of the time.
• Comparison to Real Humans: In similarity evaluations across 11 languages representing 2,222 human responses, the synthetic output proved statistically indistinguishable from actual human recordings. In these tests:
  • 45.5% of evaluators preferred the MAI-Voice-2 generated speech.
  • 44.0% of evaluators preferred the real human recording.
  • 10.5% of evaluators rated the match as a tie.

This “coin flip” distribution demonstrates that listeners cannot reliably distinguish MAI-Voice-2’s synthetic output from genuine human speech.

Safety Safeguards and the Azure Consent Workflow

Because zero-shot cloning makes voice replication simple, the technology carries high risks for misuse, including phone scams and unauthorized deepfakes. To prevent abuse, Microsoft enforces system-level consent. No unlicensed or unauthorized voice cloning is permitted in production environments.

Deploying a custom cloned voice for application development requires completing a multi-step verification process:

1. Limited Access Application: Developers must apply for approval via the Azure AI Custom Neural Voice Limited Access Review.
2. Consent Recording Upload: The user must upload a specific audio file of the target speaker reading a standard consent statement, explicitly authorizing the creation of their digital voice profile.
3. Acoustic Verification: The Azure system compares the acoustic properties of the consent recording against the reference audio clip.
4. Profile Activation: Only after successful verification does the platform unlock the Personal Voice APIs, allowing developers to generate speech using the custom voice profile.

Accessing the Model and the MAI Sandbox

For experimentation, Microsoft has launched the web-based MAI Playground. While the sandbox is free to use, it is currently geofenced and restricted to users located in the United States. Users can log in using a standard Microsoft or Microsoft 365 account without providing a credit card, paying any fees, or configuring developer APIs.

In addition to MAI-Voice-2, the playground provides access to other models in Microsoft’s in-house AI lineup, such as:

• MAI-Image-2.5: A text-to-image generator.
• MAI-Transcribe-1.5: A high-speed, speech-to-text transcription model that supports 43 languages on the FLEURS benchmark and can transcribe an hour of audio in under 15 seconds.
• DuoAI: An experimental multi-agent demo that chains transcription, reasoning, and speech synthesis together.

Developers looking to integrate MAI-Voice-2 into production software must use Microsoft Foundry or Azure Speech, where pricing for the premium TTS model starts at $22 per million synthesized characters. When integrating via Azure Speech, the model can be called using Speech Synthesis Markup Language (SSML):

<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis" xml:lang="en-US">
    <voice name="en-US-Harper:MAI-Voice-2">
        This is a synthesized speech request using Microsoft's MAI-Voice-2 engine on Azure.
    </voice>
</speak>

Microsoft is currently integrating MAI-Voice-2 into Dynamics 365 Contact Center and VS Code. In the coming months, these in-house AI models are slated for deployment across standard Microsoft productivity applications, including PowerPoint, Word, Teams, and Copilot.