On April 2, 2026, Microsoft AI announced three new foundational models under the MAI family: MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2. The release is notable not for its timing — it was expected — but for its technical specifics. Each model targets a concrete performance gap against existing alternatives.
This post breaks down what each model does, where it performs well, and how to think about integrating them into applications via Microsoft Foundry.
MAI-Transcribe-1: Speech Recognition
MAI-Transcribe-1 is a speech-to-text model supporting 25 languages. The key metrics:
- 2.5x faster than Microsoft's own Azure Fast offering for speech transcription
- 3.8% average Word Error Rate (WER) across the top 25 languages on the FLEURS benchmark
- Outperforms OpenAI's Whisper-large-v3 on all 25 supported languages
- Engineered for noisy real-world environments, not just clean studio audio
For applications that currently use Whisper or Azure Speech-to-Text, this is worth a direct benchmark. The WER improvement isn't marginal — dropping from ~5-6% WER to 3.8% on real-world audio means meaningfully fewer transcription errors in customer calls, meeting transcripts, and voice interfaces.
The pricing model is per-hour: $0.36/hour of transcribed audio. At scale, that's competitive with existing managed transcription services.
# Example: Transcribe audio using MAI-Transcribe-1 via Azure AI Foundry
from azure.ai.inference import ChatCompletionsClient
from azure.core.credentials import AzureKeyCredential
import base64
client = ChatCompletionsClient(
endpoint="https://YOUR_FOUNDRY_ENDPOINT.services.ai.azure.com/models",
credential=AzureKeyCredential("YOUR_KEY"),
)
with open("meeting_audio.mp3", "rb") as f:
audio_b64 = base64.b64encode(f.read()).decode()
response = client.complete(
model="MAI-Transcribe-1",
messages=[
{
"role": "user",
"content": [
{"type": "audio_url", "audio_url": {"url": f"data:audio/mpeg;base64,{audio_b64}"}},
{"type": "text", "text": "Transcribe this audio accurately."}
]
}
]
)
print(response.choices[0].message.content)MAI-Voice-1: Text-to-Speech
MAI-Voice-1 is a text-to-speech model with two capabilities that stand out from current TTS offerings:
- Generates 60 seconds of natural-sounding audio in a single second — real-time factor well below 1.0
- Supports custom voice creation from a few seconds of reference audio
- Preserves speaker identity across long-form content without drift
The real-time generation speed matters for interactive applications. Latency in TTS is often the bottleneck in voice assistants and phone IVR systems — if synthesis takes 2-3 seconds to produce 10 seconds of audio, the conversation feels unnatural. MAI-Voice-1's throughput removes that constraint.
Custom voice cloning from minimal reference audio is the other interesting capability. Building a consistent AI voice persona for a product — or preserving a specific speaker voice for accessibility applications — previously required multi-hour audio recording sessions. A few seconds of reference audio changes that equation.
Pricing: $22 per 1 million characters of text input.
# Example: Generate speech with a custom voice using MAI-Voice-1
import requests
payload = {
"model": "MAI-Voice-1",
"input": "Welcome to the quarterly review. Let's start with the key metrics.",
"voice": "custom",
"voice_reference_url": "https://your-storage.blob.core.windows.net/voices/speaker_reference.wav",
"response_format": "mp3"
}
headers = {
"Authorization": "Bearer YOUR_FOUNDRY_KEY",
"Content-Type": "application/json"
}
response = requests.post(
"https://YOUR_FOUNDRY_ENDPOINT.services.ai.azure.com/v1/audio/speech",
json=payload,
headers=headers
)
with open("output.mp3", "wb") as f:
f.write(response.content)MAI-Image-2: Image Generation
MAI-Image-2 is the image generation model in the set:
- Debuted in the top-3 on Arena.ai's image generation leaderboard
- 2x faster generation compared to its predecessor on Foundry and Copilot
- Available for both text-to-image and image-to-image workflows
Pricing is split: $5 per 1 million tokens for text input, and $33 per 1 million tokens for image output. For high-volume generation workloads (product visuals, marketing assets, UI mockups), the per-token image pricing needs to be modeled against actual generation volumes.
The Arena.ai leaderboard performance positions MAI-Image-2 alongside Midjourney and Stable Diffusion 3 in quality, which makes it relevant for teams already on Azure who want to avoid a separate image generation API dependency.
How This Fits Into Application Architecture
The three models together cover the core multimodal pipeline: audio input → text → audio output, with image generation as a parallel branch. For applications that need all three — a voice-enabled assistant that can also generate visual content — having them under a single Azure AI Foundry endpoint simplifies auth, billing, and rate limit management.
The more practical short-term use case is selective adoption. Teams that currently use a third-party transcription service, a separate TTS API, and a separate image generation model can consolidate onto Foundry, reducing the number of external API dependencies and likely the total cost at scale.
What to Watch
Microsoft's MAI models are built and maintained by the MAI Superintelligence team led by Mustafa Suleyman. This is distinct from the existing Azure Cognitive Services and Azure OpenAI teams — the MAI team is building first-party models that compete with, rather than wrap, OpenAI's offerings. That separation is relevant for vendor strategy: organizations concerned about single-vendor dependence on OpenAI now have a Microsoft-native alternative with clear benchmark comparisons.
The models are currently available in Azure AI Foundry (formerly Azure AI Studio). Integration with existing Azure infrastructure — identity, networking, monitoring — works through standard Azure SDK patterns.