Automated Accessibility Enhancer
An open-source tool that fixes common accessibility problems in PowerPoint files. It puts slide titles first in the reading order, adds alt text to images that don't have any, and creates a report of every change.
Updates
-
Attending EDUCAUSE 2026 in Denver
I'll be at the conference in person as an EDUCAUSE Pipeline scholar. If you'd like to meet up and talk about accessible course materials, send me an email.
-
Trying alt text from Claude on Amazon Bedrock
A hosted model would describe images instead of one running locally, so no GPU would be needed. This is on the aws-deployment branch and not merged yet. Whether to switch is tracked in #97.
-
Trying an AWS deployment with Terraform
On the same branch, the pipeline runs on AWS Lambda: upload a deck to S3 and the fixed file and report are written back to S3. It is not merged yet, and the web app still runs locally.
-
Reading order and alt text bug fixes
Fixed four bugs, including one where alt text written by a person could be overwritten if it matched the shape name. The test suite is now at 123 tests.
-
Web version
Upload files in the browser, track progress, read the report for each file, and download the fixed deck. The command-line script still works for folders.
-
EDUCAUSE 2025 Online poster session
Presented Building Institutional Community of Practice around IT Accessibility with Lalitha Subramanian.
How It Works
The tool is written in Python and uses python-pptx to edit files directly. Files can come in three ways, and all of them go through the same pipeline.
Input
- Web uploadFlask site
- Folderz_reorder.py
- S3 bucketAWS Lambda, not merged yet
Pipeline
- ConvertOlder .ppt files are converted to .pptx with LibreOffice
- Fix reading orderShapes named "Title" are moved to the front
- Add alt textImages without alt text are described by a vision model running locally
Output
- Fixed .pptxReady to download
- ReportEvery change, slide by slide
Example: one slide, before and after
Screen readers read shapes in the order they were added to the slide, not top to bottom. The numbers show the order a screen reader would read each shape. Each slide below is marked up as a list in that same order.
Before
After
Current Limitations
- Title detection can miss unusual layouts. When a slide has no real title placeholder, a small trained model guesses the title from size, position and wording. A title that is an image is not found.
- Generated alt text needs to be reviewed. Models can get names, dates, and text inside diagrams wrong. The report lists every description so they can be checked.
- Technical diagrams are hard to describe. Earlier tests with free tools like alternat did not produce usable alt text for engineering figures.
- The alt-text model is large. It is about 4.4 GB and slow without a GPU. A hosted model is being considered (#97).
- There are no user accounts. Files are kept separate with a signed browser cookie, which is not real security.
- Files are processed one at a time, and progress is stored in memory. If the server restarts during a run, that file stays marked as processing.
What's Next
- University sign-in
- Configurable rules for headings and slide structure
- Moving the web app to AWS
See Project Progress for milestones and open decisions. Feedback on the design or code is welcome.
Talks and Links
Contact
Tanya Nair, Advancing Accessibility for Engineering Education intern at the University of Washington