Microsoft AI Launches MAI-Transcribe-2

Microsoft recently launched MAI-Transcribe-2, an automated speech recognition (ASR) model designed to turn raw audio into structured, clean text. It is already available via API endpoints on Microsoft Foundry, MAI Playground, and Open Router. MAI-Transcribe-2 on Azure Speech is currently in public preview and is only available in the East US, West US, South-east Asia, and North Europe regions.

To drive initial adoption, Microsoft is offering a promotional rate of $0.10 per hour of audio that needs to be processed through the end of the year. That is a significant drop from the $0.36 rate of its predecessor, which launched just five months ago.

Audio Ingestion and Pipeline Architecture

From an integration perspective, the model runs as a structured pipeline. It takes unconditioned audio inputs of max 300 MB and serializes them into standardized, clean text.

[Raw Audio Input]
       |
       v
[Automatic Language Identification]
       |
       v
[Code-Switching Engine]
       |
       v
[Keyword Biasing Filter]
       |
       v
[Speaker Diarization Engine]
       |
       v
[Formatting Component]
       |
       v
[Structured Text Output]

The pipeline handles ingestion and processing in modular stages:

Language Identification: The pipeline automatically identifies the spoken language without requiring manual API parameters. It currently supports 60 languages, up from 43 in version 1.5 and 25 in version 1.0.
Code-Switching: The model smoothly processes multilingual inputs where speakers switch languages mid-sentence, including dialect variations like Spanglish and Hinglish.
Keyword Biasing: Developers can supply a custom list of terms, such as product names, medical terminology, or industry jargon. The model's filter matches phonetic inputs against this dictionary to maintain spelling accuracy.
Speaker Diarization: For environments with multiple speakers, the model splits the audio stream and tags distinct voices, preventing a single block of unstructured text.
Timestamps: Every transcribed word includes precise Unix-style timestamps, making it easy to align text with video frames.
Flexible Formatting: API configuration flags let users customize the style of the outputs. Setting the format to verbatim preserves filler words and pauses for legal compliance, while setting it to clean strips them out for standard reading.

Performance Benchmarks and Computing Efficiency

Third-party tests and standard benchmarks show strong performance gains in accuracy, throughput, and latency.

1. FLEURS Multilingual Benchmark

MAI-Transcribe-2 took the top spot across all 60 languages tested, posting an average Word Error Rate (WER) of 5.2%. While version 1.5 achieved a lower average WER of 3.7%, that benchmark group was limited to just 43 languages.

2. Artificial Analysis Leaderboard

On the independent Artificial Analysis leaderboard, which measures performance on corporate meetings, financial calls, and political speeches, the model ranks second in overall accuracy. It is positioned directly on the "Pareto Frontier" for latency and precision. In other words, no competitor is faster without sacrificing accuracy, and none are more accurate without adding latency.

3. Throughput and Processing Speed

In batch tests run by Artificial Analysis, MAI-Transcribe-2 finished assignments significantly faster than its nearest competitors:

5x faster than Google’s Gemini 3.5 Transcribe
7x faster than ElevenLabs’ Scribe v2
10x faster than OpenAI’s GPT-Transcribe

One important caveat is that Alibaba's proprietary model still holds the lead on raw accuracy metrics.

Compute Economics and Enterprise Integration

To move quickly and bypass standard corporate bureaucracy, Microsoft ran development through a small, ten-person engineering team, offloading data acquisition to a separate department. According to Microsoft AI CEO Mustafa Suleyman, this lean approach yielded a model that runs at half the hardware compute cost of comparable high-end options.

This release marks an important shift in Microsoft's broader AI strategy. Following its $13 billion investment in OpenAI, Microsoft renegotiated its partnerships in late 2025 and early 2026 to ensure more independence for its internal model development. The company is actively phasing out third-party OpenAI dependencies in favor of native models like MAI-Transcribe-2. Moving forward, expect to see this model powering transcription services directly inside Teams, Word, Excel, and Dynamics 365.