Microsoft has made MAI-Transcribe-2 generally available through Microsoft Foundry, moving its speech-to-text model beyond the public preview stage and allowing developers to use it in production applications. The announcement was made on November 12, 2026, with the model priced at $0.10 per hour of audio under a limited-time promotional offer. The release gives businesses and developers access to Microsoft’s transcription technology for applications involving meetings, customer support, healthcare documentation and multilingual audio processing.
Developed by Microsoft’s AI division, MAI-Transcribe-2 is designed to convert spoken audio into text while identifying individual speakers, adding word-level timestamps and recognising specialised terminology. It supports 60 languages and offers configurable transcription styles. Microsoft initially introduced the model in September 2026 as a public preview. Its general availability marks an important step towards wider commercial adoption through the company’s cloud AI platform.
MAI-Transcribe-2 Moves From Preview to General Availability
MAI-Transcribe-2 was introduced by Microsoft AI on September 3, 2026. At the time, developers could test the model through Microsoft Foundry and Microsoft’s MAI Playground.
The initial release highlighted the model’s accuracy, processing speed and ability to handle challenging real-world audio. Its general availability on November 12 means developers can now use the model for production workloads through Microsoft Foundry.
The model is part of Microsoft’s broader effort to develop and distribute its own AI technologies rather than relying exclusively on third-party models. Microsoft’s MAI portfolio includes systems for speech recognition, voice generation, image generation and other AI applications.
For businesses, the availability of MAI-Transcribe-2 provides another option for building transcription services directly into their applications and workflows.
Key Features of Microsoft’s MAI-Transcribe-2
The model includes several features designed to improve transcription accuracy and make generated transcripts more useful for professional applications.
1. Support for 60 languages
MAI-Transcribe-2 supports 60 languages, allowing developers to process audio from multilingual environments using a single transcription model.
This capability could be useful for international businesses, customer support operations, educational platforms and organisations that handle recordings in multiple languages.
Automatic language recognition also helps reduce the need for users to manually specify the language before processing audio.
2. Speaker diarization
Speaker diarization allows the model to identify different speakers in an audio recording and associate their words with the corresponding speaker.
For example, in a business meeting involving several participants, the system can separate the conversation into speaker-labelled sections.
This is particularly useful for meeting summaries, interview transcripts, call-centre recordings and legal documentation, where identifying who said what is important.
3. Word-level timestamps
The model provides timestamps at the individual-word level rather than attaching timing information only to larger sections of a transcript.
This allows developers to connect written text with the corresponding moment in an audio or video recording.
The feature can support video subtitle generation, audio editing, searchable recordings and applications that allow users to jump directly to a particular word or statement.
4. Keyword biasing
MAI-Transcribe-2 includes keyword biasing, which helps the system recognise specific words and terminology supplied by developers.
This is important in industries where recordings contain technical vocabulary, company names, product names, medical terminology or specialised abbreviations.
For example, a healthcare application could provide relevant medical terms to improve recognition, while a financial services platform could supply company names and industry-specific vocabulary.
5. Configurable transcription styles
Developers can select different transcription styles depending on the intended use.
| Transcription style | How it works | Potential application |
|---|---|---|
| Verbatim | Preserves filler words, false starts and spoken expressions | Legal records and compliance |
| Clean | Produces a more readable transcript by removing unnecessary speech elements | Meeting notes and documentation |
The ability to choose between these styles makes the model suitable for applications that require either a close representation of the original speech or a cleaner version for everyday reading.
MAI-Transcribe-2 Pricing: $0.10 per Audio Hour
One of the most notable aspects of the release is its pricing.
Microsoft is offering MAI-Transcribe-2 at $0.10 per hour of audio as a limited-time promotional price through December 31, 2026. The standard listed price is $0.36 per audio hour.
The promotional pricing makes the model an inexpensive option for developers processing large volumes of recorded audio.
| Pricing detail | Amount |
|---|---|
| Promotional price | $0.10 per audio hour |
| Standard price | $0.36 per audio hour |
| Promotional period | Through December 31, 2026 |
| Cost for 100 audio hours | $10 |
| Cost for 1,000 audio hours | $100 |
At the promotional rate, processing 1,000 hours of audio would cost approximately $100, before any applicable additional service charges or taxes.
The difference between the promotional and standard rates is significant. At $0.36 per hour, the same 1,000 hours would cost $360.
Developers planning long-term applications will therefore need to consider the standard pricing when estimating operating costs beyond the promotional period.
Microsoft Claims Strong Accuracy and Faster Processing
Microsoft says MAI-Transcribe-2 delivers strong speech-recognition performance across multiple languages and challenging audio conditions.
In its September announcement, the company reported an average Word Error Rate (WER) of 5.2% on the FLEURS benchmark across 60 languages. Word Error Rate measures how frequently a transcription system makes mistakes involving substitutions, deletions and insertions.
Microsoft also highlighted the model’s performance in Artificial Analysis evaluations, describing it as a strong combination of accuracy and processing latency.
The company said MAI-Transcribe-2 can process long recordings up to 10 times faster than leading competitors in certain evaluations.
These are Microsoft-reported benchmark results rather than guarantees of performance for every recording. Actual accuracy can vary depending on background noise, microphone quality, accents, overlapping speech and the subject matter being discussed.
Where Businesses Can Use MAI-Transcribe-2
The model’s combination of speaker identification, timestamps, multilingual support and keyword recognition makes it relevant to several industries.
Healthcare: Hospitals and healthcare software developers could use transcription systems to convert recorded consultations and clinical discussions into searchable text. Such applications would still need appropriate privacy, security and human-review safeguards.
Legal services: Legal professionals could use verbatim transcription to prepare searchable records of interviews, hearings and recorded conversations.
Customer support: Businesses could analyse call recordings, identify speakers and review conversations for quality assurance and customer-service improvements.
Education: Schools, universities and online learning platforms could use the model to generate transcripts and improve accessibility for recorded lectures.
Media and entertainment: Production companies could use word-level timestamps to support subtitle generation, editing and searching through large collections of audio and video content.
The Bigger Picture
Microsoft’s decision to make MAI-Transcribe-2 generally available strengthens its position in the growing market for AI-powered speech recognition. Transcription is becoming an important component of business software, voice assistants, customer-service systems and AI agents that need to understand spoken instructions.
The pricing strategy is also significant. At $0.10 per audio hour through the end of 2026, Microsoft is making the model attractive to developers that process substantial amounts of audio. A combination of low pricing, multilingual support and advanced transcription features could help the company compete with established speech-to-text providers.
However, production adoption will depend on more than price and benchmark performance. Developers will need to evaluate transcription accuracy across their own languages, accents and recording conditions. They will also need to consider data protection, regional availability and the standard price after the promotional period ends.
Looking Ahead
With MAI-Transcribe-2 now generally available in Microsoft Foundry, developers can begin integrating the model into production applications and testing its performance against their existing transcription systems. The promotional price of $0.10 per audio hour provides an opportunity to evaluate large-scale workloads at a relatively low cost before the offer expires at the end of 2026.
The release also demonstrates Microsoft’s broader ambition to expand its internally developed AI model portfolio across different business applications. As demand grows for voice-enabled software, multilingual assistants and automated documentation, speech recognition could become an increasingly important part of enterprise AI infrastructure. The model’s long-term adoption will depend on its real-world accuracy, reliability and pricing competitiveness.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
