Our AI Engineer, Safra Soymat, walks us through her recent work on the UDL Classroom, exploring how a Mac mini paired with offline speech recognition can create truly secure and accessible learning resources.

Recently, I’ve been testing something exciting: using a Mac mini to run speech recognition completely offline.

Why does this matter?

In education, privacy and accessibility often clash. We want students to have transcripts, captions, and study tools. However, we don’t want their voices or lecture recordings sent to third-party servers. This is especially important in sensitive classrooms with specific accessibility needs.

So, I set out to answer a straightforward question: can we make this work locally, on a Mac mini, with Apple Silicon?
‍

What I’ve Been Exploring

As part of our project, I’ve looked into a few areas:

• Running speech recognition locally: I tested small and medium Core ML–converted models on the Mac mini to find the right balance between speed and accuracy.
• Performance trade-offs: I benchmarked 1, 2, 4, and 7-hour lecture videos to see how transcription time, accuracy, and temperature held up.
• Privacy first workflows: I ensured the pipeline accepts MP4 video files and outputs transcripts in TXT and SRT format without using the cloud.
• Hardware limits: I pushed the Mac Mini to check CPU, GPU, and Neural Engine utilization while monitoring the maximum temperature to prevent overheating.
• UDL use cases: I explored how to transform the transcripts into captions, searchable notes, or study aids that meet various student needs.
  ‍

The fun part? It’s not just theoretical. We’ve already built a working prototype. Feed in a video, get a transcript back, and know that the entire process happened locally on the Mac mini in the classroom.
‍

Why Offline First?

Most speech-to-text tools rely on the cloud. They’re useful, but they also:

• send sensitive data off the device,
• depend on strong Wi-Fi (which isn’t always reliable in colleges and universities), and
• take control away from educators and learners.
  ‍

If we keep the processing in the classroom, on the device, we can improve the situation. We protect the data, respect institutional policies, and still provide high-quality AI support to students.
‍

The Mac mini in Action

With the Neural Engine speeding up inference, we’ve shown that the Mac mini can handle long lectures, produce accurate transcripts, and do so without overloading. This isn’t just about running a model. It’s about ensuring it works under real classroom conditions.
‍

Why This Matters for UDL

Universal Design for Learning (UDL) focuses on making education flexible, inclusive, and accessible. AI transcription plays a significant role here:

• Students who are hard of hearing receive captions.
• Students who learn best by reading get full transcripts.
• Neurodiverse learners access structured content they can revisit.
• Everyone benefits from searchable, organized notes.

When we run this locally, all these benefits come without sacrificing privacy.
‍

Looking Ahead

This exploration has shown me that compact, affordable devices like the Mac mini can be privacy-first AI tools in education. The next step is to scale this approach; more classrooms, more formats, and more adaptive features that deliver what students need, when they need it.

For me, that’s the exciting part. It’s not just about the technology, but about creating a future where AI supports every learner while maintaining trust.