MAI Transcribe-1.5 - A speech-to-text model launched by Microsoft MAI

What is MAI Transcribe-1.5

MAI-Transcribe-1.5 is a speech-to-text model independently developed by the Microsoft AI team.speech-to-textIt supports 43 languages and features context-aware keyword biasing. The model achieves the industry's lowest word error rate (WER 4.86%) on the FLEURS benchmark, and is launched specifically for enterprise-grade production scenarios such as video subtitles, meeting transcription, and call analysis.

Main features of MAI Transcribe-1.5

• High-accuracy transcription in 43 languages: Covers 43 languages including English, Chinese, Japanese, Hindi, and Arabic, with automatic language identification.
• Keyword/entity biasing: Up to 200 domain-specific terms (such as personal names, product names, and medical terminology) can be injected, using context to intelligently determine whether to apply the bias rather than forcing a match.
• Robustness in noisy environments: Optimized for real-world background noise and variable-quality audio while maintaining high accuracy.
• Ultra-fast processing of long audio: 1 hour of audio takes about 15 minutes to transcribe, up to 5 times faster than the previous generation.
• Adaptability to industry scenarios: Built-in understanding of terminology in fields such as healthcare, customer service, and finance, ready to use out of the box.

Technical principles of MAI Transcribe-1.5

• Unified multilingual modeling: The model is jointly trained on massive speech data in 43 languages, covering major languages and including low-resource languages such as Assamese, Gujarati, and Kannada. Through shared representation learning, it achieves cross-lingual transfer, ensuring stability across different accents and dialects.
• Context-aware keyword biasing mechanism: Unlike traditional forced replacement, MAI-Transcribe-1.5 incorporates user-provided domain vocabulary into the decoding process as soft prompts. The model combines acoustic features and semantic context to dynamically determine when to activate the biasing strategy. On the FLEURS multilingual benchmark, it can further reduce WER by 30% while avoiding harm to general vocabulary.
• Long-audio segmentation and streaming optimization: For long-duration audio such as meetings and podcasts, the model uses improved segmentation and caching mechanisms to reduce redundant computation and memory usage, significantly lowering end-to-end latency while maintaining semantic coherence across segments.

How to use MAI Transcribe-1.5

• Azure Speech SDK: Integrate the SDK into an application and call the MAI-Transcribe-1.5 model endpoint, supporting WAV/MP3/FLAC formats (single file up to 300 MB or 2 hours).
• REST API: Send audio streams or files directly via HTTP requests to obtain transcription results in JSON format.
• MAI Playground: On the Microsoft Mai playground official website https://playground.microsoft.ai/的交互式沙盒中上传音频，即时体验效果。
• Microsoft Foundry: Access via the Azure Speech service, billed at $0.36/hour of audio, with no need to deploy the model.

Core advantages of MAI Transcribe-1.5

• Industry-leading accuracy: FLEURS 43-language average WER is 4.86%, lower than Elevenlabs Scribe v2 (5.53%), OpenAI Transcribe (5.73%), and Google Gemini Flash Lite (5.63%).
• Doubled language coverage: Compared with v1 25 languages, 18 languages have been added, making it more suitable for global products.
• Zero errors on domain vocabulary: Through keyword biasing, it accurately transcribes proprietary terms, abbreviations, and drug names within enterprises.
• Balances cost and speed: The pricing of $0.36/hour combined with 5x faster long audio processing makes the cost-effectiveness even more outstanding.

Project address of MAI Transcribe-1.5

• Project official website：https://microsoft.ai/models/mai-transcribe-1-5/
• Technical paper：https://microsoft.ai/pdf/MAI-Transcribe-1.5-Model-Card.PDF

Comparison of MAI Transcribe-1.5 with similar competing products

Application scenarios of MAI Transcribe-1.5

• Video subtitles and content localization: Automatically generate high-accuracy subtitles in 43 languages for global video platforms, reducing localization costs.
• Meeting and interview transcription: Quickly convert multilingual meeting recordings into searchable text, with 1 hour of audio completed in about 15 minutes.
• Customer service call analysis: Accurately recognize professional terms such as drug names and product models to support intelligent quality inspection and sentiment analysis.
• Medical dictation records: Automatically transcribe anatomical and pharmaceutical terms in doctors' rounds and surgical records, improving the efficiency of medical record entry.
• Accessibility assistive tools: Provide real-time speech-to-text services for people with hearing impairments, supporting clear recognition in noisy environments.