MAI-Transcribe-2 - Microsoft's most powerful AI speech-to-text model

What is MAI-Transcribe-2

MAI-Transcribe-2 is Microsoft's most powerfulAI speech-to-textmodel, leading across the board in accuracy, speed, and cost. The model supports 60 languages, with an average word error rate of only 5.2% on the FLEURS test set. The model adds speaker diarization and word-level timestamps, offers two transcription styles, verbatim (retaining slips of the tongue) and clean (removing filler words), and supports automatic language identification and language switching.

MAI-Transcribe-2's main features

• Multilingual speech transcription: Supports 60 languages and can automatically detect the language of a recording without specifying the language in advance.
• Speaker diarization: Distinguishes different speakers within a single recording and attributes the transcribed text to the corresponding speaker.
• Word-level timestamps: Marks the precise time position of each word, making subtitle alignment, audio search, and editing easier.
• Configurable transcription styles: verbatim mode retains filler words and slips of the tongue, while clean mode removes filler words to generate easy-to-read text.
• Automatic language identification and switching: Supports conversations that naturally mix languages, automatically detecting and adapting to language switches.
• Keyword biasing: Allows presetting a list of domain terms to improve recognition accuracy for specific vocabulary.
• Transcription in noisy environments: Has the ability to maintain high transcription accuracy in background noise.

MAI-Transcribe-2's technical principles

• End-to-end unified architecture: Uses a single neural network to directly map acoustic features to text sequences, integrating speech recognition, speaker diarization, language detection, and other tasks into one model, avoiding the error accumulation of traditional cascaded systems and achieving optimal speed and accuracy simultaneously.
• Joint multilingual training: Based on joint training with massive speech-text data in 60 languages, the model learns cross-lingually shared acoustic representations and language patterns, enabling it not only to cover major languages but also to maintain high accuracy for low-resource languages, with real-time detection capability for natural language switching.
• Context-aware decoding: During decoding, the model not only predicts based on the current acoustic frame but also performs joint reasoning with global contextual semantics and external keyword biasing, dynamically adjusting the probability distribution of output words, improving the recognition accuracy of technical terms, and supporting controllable generation in both verbatim and clean styles.

How to use MAI-Transcribe-2

• Create an Azure resource: Create an AI Speech service resource in the Azure portal to obtain an API key and regional endpoint address.
• Access the Foundry preview: Through the Microsoft Foundry public preview entry https://ai.azure.com/catalog/models/MAI-Transcribe-2进入模型调用界面。
• Configure model parameters: In the API request, specify model: "MAI-Transcribe-2", and enable speaker diarization and word-level timestamps as needed.
• Upload audio for transcription: Submit audio files in formats such as WAV/MP3/MP4, and the model automatically detects the language and performs transcription.
• Get structured results: Receive the returned JSON, including the full text, word-level timestamps, speaker labels, and the automatically detected language code.

MAI-Transcribe-2's core advantages

• World-leading accuracy: FLEURS averages only 5.2% word error rate across 60 languages, with Chinese as low as 4.5%, comprehensively outperforming Gemini 3.1 Pro and Whisper v3-Large.
• Fastest in the industry: Processing speed reaches 403.6x real-time, completing 1 hour of audio in about 9 seconds, 10x faster than GPT-Transcribe.
• Lowest cost in the industry: Limited-time pricing of $0.10/hour, about 72% cheaper than the previous generation.
• Native feature integration: Built-in speaker diarization, word-level timestamps, automatic language recognition and switching, and verbatim/clean dual-style transcription, with no additional post-processing required.
• Seamless multilingual coverage: Supports 60 languages, can automatically detect and adapt to naturally mixed-language conversations, with no need to specify the language in advance.

Comparison of MAI-Transcribe-2 with similar competitors

Application scenarios of MAI-Transcribe-2

• Call center quality inspection and analysis: Use speaker diarization to distinguish agents from customers, and combine word-level timestamps to measure script ratios and precisely locate key moments.
• Intelligent recording of meetings and interviews: Convert multi-person recordings into structured text with speaker labels, automatically attributing action items and decision sources.
• Real-time dictation for voice agents: As a low-latency auditory input layer, it pairs with MAI-Voice-2 Flash to enable a real-time speech-to-speech conversation loop.
• Media subtitles and localization: Word-level timestamps greatly reduce subtitle alignment costs, and coverage of 60 languages supports global content production.
• Compliance, legal, and regulatory evidence preservation: Generate precise records with speaker attribution and timestamps, meeting the audit and forensic requirements of industries such as finance and healthcare.