MAI-Transcribe-2 Operational

MAI-Transcribe-2 is Microsoft's most powerful AI speech-to-text model, leading across accuracy, speed, and cost. It supports 60 languages and achieves an average word error rate of just 5.2% on the FLEURS test set. The model adds speaker diarization and word-level timestamps, offers verbatim (retaining speech disfluencies) and clean (removing filler words) transcription styles, and supports automatic language identification and language switching.

Multilingual speech transcription: Supports 60 languages and automatically detects the recording language without requiring a language to be specified in advance.
Speaker diarization: Distinguishes different speakers within a recording and attributes the transcribed text to the corresponding speaker.
Word-level timestamps: Assigns precise time positions to each word for subtitle alignment, audio search, and editing.
Configurable transcription styles: Verbatim mode retains filler words and speech disfluencies, while clean mode removes filler words to produce more readable text.
Automatic language identification and switching: Supports conversations that naturally mix languages, automatically detecting and adapting to language changes.
Keyword biasing: Allows predefined lists of domain-specific terms to improve recognition accuracy for particular vocabulary.
Transcription in noisy environments: Maintains high transcription accuracy in environments with background noise.

Unified end-to-end 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. This avoids error accumulation in traditional cascaded systems while optimizing both speed and accuracy.
Joint multilingual training: Trained jointly on extensive speech-text data across 60 languages, enabling the model to learn shared cross-lingual acoustic representations and language patterns. It covers major languages, maintains high accuracy for low-resource languages, and provides real-time detection of natural language switching.
Context-aware decoding: During decoding, the model predicts using not only the current acoustic frames but also global contextual semantics and external keyword biasing for joint inference. It dynamically adjusts the probability distribution of output words, improving recognition of specialized terminology and supporting controllable generation in both verbatim and clean styles.

Create an Azure resource: Create an AI Speech service resource in the Azure portal and obtain an API key and regional endpoint URL.
Access the Foundry preview: Enter the model interface through the Microsoft Foundry public preview at https://ai.azure.com/catalog/models/MAI-Transcribe-2.
Configure model parameters: Specify model: "MAI-Transcribe-2" in the API request, and enable speaker diarization and word-level timestamps as needed.
Upload audio for transcription: Submit audio files in formats such as WAV/MP3/MP4. The model automatically detects the language and transcribes the audio.
Receive structured results: Receive the returned JSON containing the complete text, word-level timestamps, speaker labels, and automatically detected language code.

Industry-leading accuracy: Achieves an average word error rate of just 5.2% across 60 FLEURS languages and as low as 4.5% for Chinese, outperforming Gemini 3.1 Pro and Whisper v3-Large overall.
Fastest speed in the industry: Processes audio at 403.6 times real time, completing approximately 1 hour of audio in 9 seconds, 10 times faster than GPT-Transcribe.
Lowest cost in the industry: Limited-time pricing of $0.10 per hour, approximately 72% lower than the previous generation.
Native feature integration: Includes speaker diarization, word-level timestamps, automatic language identification and switching, and verbatim/clean dual-style transcription without additional post-processing.
Seamless multilingual coverage: Supports 60 languages and automatically detects and adapts to conversations that naturally mix languages, without requiring a language to be specified in advance.

Call center quality assurance and analysis: Uses speaker diarization to distinguish agents from customers, combined with word-level timestamps to measure talk-time ratios and precisely locate key moments.
Intelligent meeting and interview records: Converts multi-speaker recordings into structured text with speaker labels, automatically attributing action items and decisions to their sources.
Real-time dictation for voice agents: Serves as a low-latency auditory input layer and, when paired with MAI-Voice-2 Flash, enables a real-time speech-to-speech conversational loop.
Media subtitling and localization: Word-level timestamps significantly reduce subtitle alignment costs, while coverage of 60 languages supports global content production.
Compliance, legal, and regulatory records: Generates precise records with speaker attribution and timestamps to meet audit and evidentiary requirements in industries such as finance and healthcare.