Microsoft has launched a new version of its speech-to-text model called MAI-Transcribe-1.5. It improves on accuracy and speed, making it one of the fastest and most precise models available.

The model processes audio in 43 languages, up from 25 in the previous version. This includes many South Asian languages like Bengali, Tamil, and Telugu, as well as European languages such as Ukrainian and Greek. This broad coverage lets companies handle diverse audio without switching models.

Accuracy is a key highlight. MAI-Transcribe-1.5 achieves a 2.4% word error rate on the Artificial Analysis benchmark, placing it third among top speech-to-text models. On the FLEURS benchmark, it holds best-in-class accuracy across all 43 languages.

What sets this model apart is its speed. It can transcribe audio about 276 times faster than real time. That means it can turn an hour of speech into text in less than 15 seconds. This speed outpaces all other top-accuracy models by a wide margin.

Why Speed and Accuracy Matter

In speech recognition, speed and accuracy often compete. Models that are very accurate tend to be slow, making them less useful for live or high-volume tasks. Faster models usually sacrifice accuracy. MAI-Transcribe-1.5 changes that dynamic.

This model sits on what’s called the accuracy-speed Pareto frontier. It offers the best accuracy possible without sacrificing speed. For businesses, this means live captions, meeting transcriptions, and voice assistants can work quickly and correctly.

For example, customer service centers can analyze calls faster without losing detail. Content creators can get fast, reliable transcripts for podcasts or videos. And healthcare providers can transcribe technical terms correctly thanks to a new feature called keyword biasing.

New Features That Improve Real-World Use

Keyword biasing helps the model recognize specific terms like names, medical jargon, or company acronyms. Without this, speech models often mishear uncommon words. This feature allows users to supply a list of important words, and the model adjusts its transcription accordingly.

Microsoft reports this feature reduces errors by 30% on complex vocabulary. It’s especially useful in fields like healthcare, legal work, and enterprise call centers where misspelled terms can cause big problems.

Another practical enhancement is automatic language identification. The model detects the spoken language without needing manual input. This is helpful for multilingual environments or when the language is unknown.

Currently, MAI-Transcribe-1.5 does not support speaker diarization, so it cannot label who is speaking. Streaming transcription is also not available yet but is planned for the future.

Cost and Availability

The model is available now through Microsoft Foundry and Azure AI services. Pricing is about $0.36 per hour of audio, which is cost-effective compared to similar high-accuracy models from other providers.

It integrates easily with Microsoft products like Teams, GitHub, Dynamics 365, and Copilot, making it a good choice for companies already in the Microsoft ecosystem. This integration simplifies deploying the model in real-world workflows.

Its ability to handle noisy or overlapping audio makes it fit for varied environments, from busy call centers to recorded meetings. Azure’s compliance with regulations like HIPAA and GDPR also makes this model suitable for industries with strict data privacy needs.

With its speed, accuracy, and new features, MAI-Transcribe-1.5 raises the bar for speech-to-text technology. It offers a powerful tool for companies that rely on fast, reliable transcription across many languages and domains.