This application provides
ten tools for accessible content
creation. Most tools require an API credential to function.
Expand a section below to configure your credentials.
Set Up
This page. Configure API credentials for every tool, and
download or remove local AI models. It requires none of its
own.
OpenRouter AI
Send prompts to a wide range of AI models. Requires an
OpenRouter API key.
Microsoft Foundry
Alternative provider for AI tools, using Microsoft AI models
via your own proxy. Requires deploying a Cloudflare Worker.
Markdown Editor
Write and preview Markdown with support for charts, diagrams,
and mathematics. Credentials are not required.
Graph Builder
Create accessible charts and graphs with data tables and alt
text. Credentials are not required.
MathPix
Convert images of handwritten or printed mathematics into
LaTeX. Requires a MathPix App ID and App Key.
Image Describer
Generate detailed alt text for images using AI. Requires an
OpenRouter API key, or download a local model below.
Chat
Hold a conversation with a wide range of AI models, filtering
them by cost, capability and provider. Use cloud models with
an OpenRouter API key or a Microsoft Foundry connection, or
download a local model below to chat with no credentials,
keeping your data on your device.
Local Chat
Chat with a local AI model running entirely in your browser.
Your data stays on your device. Credentials are not required.
You will need to download a model from the Local AI Models
section.
Ally Reporting
Query the Blackboard Ally API for course accessibility
reports. Requires an Ally API token and Client ID.
Transcribe
Turn an audio recording into a text transcript with speaker
labels and timestamps. Alpha — expect rough edges. Requires
signing in with your University account.
Captions Fixer
Correct the errors in an automatic caption file. Upload the
subtitles from a recording, review the suggested corrections,
and download a corrected file with the original timings
intact. Alpha — expect rough edges. Requires an OpenRouter API
key, or a Microsoft Foundry connection.
Word forms
Rebuild a Word form so it works with a screen reader. Choose a
.docx file, read what the checks found and the plan for the new
form, then agree to build it or stop. Alpha — expect rough
edges. Requires a Microsoft Foundry connection.
TikZ Studio
Write TikZ source and see the diagram beside it, with a text
alternative, caption and long description, then download it as
SVG or PNG or export it. Alpha — expect rough edges. Needs no
key.
API Credentials
University sign-inStatus: Not yet available
Some tools can use University of Southampton AI services.
Select the button to sign in. Signing in shows that you are a
University of Southampton member. You will be brought back
here straight afterwards.
Save anything you have typed on this page before you sign in,
because moving to Microsoft will clear it.
OpenRoutercredentialsOpenRouter: Not configured
MathPixcredentialsMathPix: Not configured
Microsoft FoundrycredentialsActiveMicrosoft Foundry: Not configured
AllycredentialsAlly: Not configured
Local AI Models
Download and manage AI models that run locally in your browser.
These models work without an API key after initial download.
Description Generation
Generate image descriptions locally — no API key or internet
connection required after initial download.
Choose which local model to use for generation in the Image
Describer tool. Download models here, then select your
preferred model when generating descriptions.
FastVLM 0.5BLocal description generation (no API key
needed)
~350 MB
Checking…
Qwen3.5-0.8BStructured local descriptions
~1,000 MB
Checking…
LFM2-VL 450MCompact local descriptions (no API key needed)
~349 MB
Checking…
Analysis Pipeline
Florence-2Captioning, object detection, OCR
~1.1 GB
Checking…
Florence-2 enhancement (AI caption and supplementary OCR) is
available when using the Image Describer tool with an
uploaded image.
Text Generation
Local text models for chat and AI tasks — no API key or
internet connection required after initial download.
LFM2 350M
ONNXFastest model — works on all hardware including
integrated GPUs
Phi 3.5 Mini
ONNXHighest quality ONNX model — requires 8 GB+ discrete
GPU
~2,300 MB
Checking…
Details & benchmarks — Phi 3.5 Mini
Strengths
Limitations
Best for
Hardware
Speed
Context safe
Rating
Llama 3.2 1B
WebLLMBest balance of speed and quality via WebLLM — works
on all GPUs
~879 MB
Checking…
Details & benchmarks — Llama 3.2 1B
Strengths
Limitations
Best for
Hardware
Speed
Context safe
Rating
Qwen 2.5 1.5B
WebLLMWebLLM mid-tier — better quality than 1B, works on
all GPUs
~1,500 MB
Checking…
Details & benchmarks — Qwen 2.5 1.5B
Strengths
Limitations
Best for
Hardware
Speed
Context safe
Rating
Llama 3.2 3B
WebLLMHighest quality with full context — works on 4 GB+
GPUs
~1,724 MB
Checking…
Details & benchmarks — Llama 3.2 3B
Strengths
Limitations
Best for
Hardware
Speed
Context safe
Rating
Text-to-Speech
Supertonic TTS
ONNXNatural-sounding speech — read alt text aloud
with near-human quality
~263 MB
Checking…
Voice Settings
This voice sends text to an external server for
processing.
When enabled, Read Aloud announces structural cues
such as "Heading level 2" and "Bulleted list with 3
items" to help you understand the layout of the
content.
Choose how code blocks are read aloud. "Summarise with
first line" announces the language, line count, and
the first non-blank line. "Summarise" announces only
the language and line count. "Read all code" reads
every character, including punctuation.
GPU Memory
No models loaded in GPU memory
Storage & Cache
Storage: checking…
Persistent
Checking cache…
Loaded (in memory)
Cached (on disk)
Not downloaded
Downloading…
OpenRouter Testing
Choose from the filtered list of AI models based on your
criteria
Select a model to view detailed information about its
capabilities, pricing, and features.
Advanced Model FiltersShowing all models
Expand to access advanced filtering options including search,
pricing, capabilities, and provider filters
Type to filter models by name, provider, or description
Clear all applied filters and show all available models
Loading models...
Enter text to process with the selected AI model
Supported: Images (JPG, PNG, WebP) up to 20MB, PDFs up to
512MB
Provide context and instructions for how the AI should
behave.
Current:
Focused (0.0)Balanced (1.0)Creative (2.0)
Controls response creativity. Lower values give focused,
consistent responses. Higher values increase variety and
creativity.
Current:
Focused (0.1)Balanced (0.5)Varied (1.0)
Adjusts how varied the model's word choices can be.
Current:
tokens
/ 4096 available
ShortMediumLong
Controls maximum length of AI responses. Longer responses
cost more tokens.
Current:
More SimilarBalancedMore Varied
Word variety:
More SimilarBalancedMore Varied
Phrase variety:
Controls how varied the AI's word and phrase choices will
be. Higher values create more diverse language, while lower
values allow more repetition. Values below 0 encourage
repetition, which can be useful for maintaining consistency.
You can provide your own custom descriptions for charts by adding
a `descriptions` object within your chart configuration:
"descriptions": {
"short": "Your concise chart description here.",
"detailed": "Your detailed HTML description here. You can use paragraphs, lists, tables and links.",
"detailedFormat": "html"
}
The detailed description is checked before it is shown. It keeps
paragraphs, headings, lists, definition lists, tables (with
captions and scope), links, and basic text
formatting. Scripts, styles, forms, images, embedded content and
event handlers are removed. Links must be http,
https, mailto or relative. Top-level
headings (h1, h2) are shown as level-4
headings, because the panel already uses level 3.
Set detailedFormat to "text" to write
plain text instead: blank lines start new paragraphs and any
angle brackets are shown literally. The default is
"html".
```mermaid
graph TB
A[Start] --> B{Is it raining?}
B -->|Yes| C[Take umbrella]
B -->|No| D[Take sunglasses]
C --> E[Go out]
D --> E
E --> F[Return home]
```
Type or paste markdown content with LaTeX math expressions,
charts, or code. Press Ctrl+Enter to render. The content will be
rendered as HTML below.
Examples demonstrate various markdown features including
tables, charts, and diagrams.
Select a CSV file from your computer or drag and drop one
into the box above. Maximum file size: 1MB.
Extract data from a chart image
An AI model reads the chart and fills the data table. Check
the numbers before building your chart, because values read
from a picture are estimates.
PNG, JPEG or WebP. You can also paste one from the clipboard
(Ctrl+V).
No image chosen
Your original image
Data Preview
You can correct any value here before you choose a chart.
Preview of your data
Developer information
Finish reason
Not reported
Request
Response
Step 2: Choose Chart Type
Your original image
Step 3: Configure Your Chart
Live Preview
Your original image
Chart Settings
Your
Chart is Ready!
Your accessible chart has been created with all the features
you need.
Your original image
MathPix Mathematics OCR
Upload Mathematics
Drop file here, or select to browse.
Supports: Images (JPEG, PNG, WebP), Documents (PDF, DOCX,
PPTX)
Quick tip: You can paste images directly from
your clipboard (Ctrl+V or Cmd+V)
Draw Mathematics
Draw your mathematical expression using your mouse or touch
input. The system will recognise handwritten mathematics.
Each Process Drawing sends one request to MathPix: about
$0.002 (MathPix list price, US
dollars).
Alternative: Drawing is only supported using
stylus / mouse / touchpad / touch or other pointing devices. For
other modes of input, please use the Upload mode to process
images instead.
Camera inactive
Starting the camera and capturing a photo are free.
Processing the photo sends one image to MathPix: about
$0.002, or up to $0.005 for
a dense page (MathPix list price, US dollars).
Position mathematics clearly in view. Captured photos will be
processed using the same options as uploaded files.
MMD Editor and Preview
Type or paste MMD content below, or drag and drop a
.mmd, .md, or
.txt file.
Use AI to improve OCR accuracy of the processed PDF
document.
Wrap long lines in the code view instead of scrolling
sideways.
Plain Markdown (MD) Output
HTML Output
HTML Preview:
LaTeX Output
DOCX Output
PowerPoint (PPTX) Output
Binary Format: PowerPoint files cannot be
previewed here. Click "Download PPTX" to save the file to
your device.
This presentation format is compatible with Microsoft
PowerPoint 2007 and later. Each page becomes a slide with
preserved mathematics and formatting.
Excel (XLSX) Output
Binary Format: Excel files cannot be
previewed here. Click "Download XLSX" to save the file to
your device.
This spreadsheet format carries the document's tables, one
sheet per table, and opens in Microsoft Excel, Google
Sheets and other spreadsheet software.
PDF (HTML Rendering) Output
Binary Format: PDF files cannot be
previewed here. Click "Download PDF" to save the file to
your device.
This PDF contains HTML-rendered mathematics suitable for
printing and sharing. Compatible with all PDF viewers.
PDF (LaTeX Rendering) Output
Binary Format: PDF files cannot be
previewed here. Click "Download PDF" to save the file to
your device.
This PDF contains LaTeX-rendered mathematics with
selectable equations. Superior typography
for academic papers and technical documents.
OCR Confidence Analysis
Visual overlay showing recognition confidence for each
detected element. Colour coding indicates confidence
levels from high (green) to low (red).
Loading confidence visualiser...
Page 1 of
1
Page Content Analysis:
Lines:
0
Mathematics:
0
Tables:
0
Handwritten:
0
Printed:
0
Export
Options
Converting sends a separate request to MathPix. MathPix
does not publish a price for conversion; check your
MathPix console for the charge.
Cancel stops this page waiting for your files. MathPix may
still charge for the request it has already received.
Full feature support. Best for users in Americas. Data
processed in US.
GDPR Compliant. Best for EU/UK users. Your
documents, results and derivatives stay in the EU. All
output formats are available.
Best for users in Asia Pacific region. Data processed in
Singapore.
Image Describer
Generate accessible alt text descriptions for images using AI.
Upload Image
Or use camera
Check the photo. Rotate or flip it if you need to, then
confirm.
Camera is not available in this browser or on this
device. Please use the file upload option above.
Position your image clearly in view. The captured photo
will be processed using the options configured above.
Detect text in this picture?
Text detection reads words and labels, such as those on a
diagram, chart or map. Choose Skip text for a painting or a
photograph of a scene.
Re-analysing discards the OCR corrections saved for this
image. They cannot be recovered.
Tip: You can use LaTeX notation here, e.g.
x^2 + y^2 = r^2
Confidence
-
Position
-
Source
-
For example: TH1.
Add Context
Educational Context
For example: Mathematics, Biology, History.
For example: differential equations, cell
division.
What should students understand after engaging with
this image?
For example: MATH2001.
Any other relevant information about how this image
is used.
Generated Description
Preparing...0s
This
description was generated by a small local model and may
contain factual errors. Please check all details carefully
against the original image before use.
Your model preference will be saved for future sessions
Visual Verification
Sends a second request to verify the description
matches the image. Uses an additional AI request.
A smaller, cheaper model often works well for
verification.
Local Generation
Generate image descriptions locally — no API key or
internet connection required after initial download.
FastVLM and LFM2-VL produce free-form prose
descriptions. Qwen3.5 produces structured sections
(title, alt text, long description, text content) but
may contain factual errors.
This model may produce factual errors. Please check all
details carefully against the original image before use.
FastVLM 0.5BLocal description generation (no API key
needed)
~350 MB
Checking…
Qwen3.5-0.8BStructured local descriptions
~1,000 MB
Checking…
LFM2-VL 450MCompact local descriptions (no API key
needed)
~349 MB
Checking…
Analysis Pipeline
Florence-2Captioning, object detection, OCR
~1.1 GB
Checking…
Florence-2 Enhancement
Run Florence-2 to enrich your description with an
AI-generated caption. This is optional and takes
approximately 30 seconds.
Storage & Cache
Storage: checking…
Persistent
Checking cache…
Loaded (in memory)
Cached (on disk)
Not downloaded
Downloading…
Ally Accessibility Reporting
Cached Data
Storage:
0 KB of 4 MB
Cached reports
No cached data. Reports will be cached automatically when you
generate them.
API Unavailable
The Ally API is currently unavailable. You can view previously
cached reports.
Browse Cached Reports
Select a cached report to view:
Module Report
Generate a detailed accessibility report for a single module.
Search for and select a module to view comprehensive accessibility
metrics, content inventory, and issue breakdown.
Type to search modules. Use arrow keys to navigate results,
Enter to select, Escape to close.
No course data has been uploaded for this institution, so
search asks the Ally service directly. It matches module
names only, not module codes, and each search can take a few
minutes while the report prepares.
Select a module above to enable this button
Preparing report...
Accessibility Statement Preview
Preview the accessibility statement that students will see for a
selected module. This shows a simplified, student-facing summary
of potential accessibility barriers.
Choose your institution and virtual learning environment. This
adjusts wording throughout the statement, such as "module lead"
or "Blackboard".
Choose which parts of the statement appear — on screen and in
every export. The module information is always included.
When enabled, exported statements include an "Update
accessibility data" button that refreshes the
accessibility-data section against the live Ally data.
This is read at export time, so it only affects statements
you export from now on. Leave it off to export a fixed
snapshot.
When enabled, statements you export open in focus mode: a
distraction-free reading view with the sidebar and table
of contents hidden. Readers can leave focus mode at any
time with the Escape key or the on-screen toggle. This is
read at export time, so it only affects statements you
export from now on.
Type at least 2 characters to search. Use arrow keys to
navigate results, Enter to select. If more matches exist,
press Down arrow at the end of the list, or Tab to the Show
more button, to load more.
No course data has been uploaded for this institution, so
search asks the Ally service directly. It matches module names
only, not module codes, and each search can take a few minutes
while the report prepares.
Select a module first to enable this buttonSelect a module first to enable this button
Preparing statement preview...
Copy as:
Module information
Accessibility Over Time
How accessibility scores have changed, for the whole institution
or for one department. This is the only view built from a periodic
export rather than a live request, because the Ally API returns
the current position only and has no way to ask for an earlier
one.
Data table
Build Query
Module accessibility scores and file type counts
Accessibility issues by category and severity
Type to search module. Use arrow keys to navigate results,
Enter to select, Escape to close.
No course data has been uploaded for this institution, so
search asks the Ally service directly. It matches module
names only, not module codes, and each search can take a few
minutes while the report prepares.
Which institution this tool is looking at, and where its
course data comes from. These are the same controls as the
Ally card in Set Up, and a change here shows there.
For example: University of Southampton
Labels this configuration so colleagues can tell which
institution a shared file is for, and it is included when
you export one. It is stored with your Ally credentials in
Set Up: selecting Save credentials above does not store it.
Off by default. The API token is a credential: anyone who
receives a file containing it can query your Ally data as
you. Leave this unticked and let each colleague enter their
own token.
Choose a file, or drop it here
No file chosen
Choosing a file replaces the region, Client ID, API token
and proxy worker URL above, so you are asked to confirm
first. Course data and cached reports are never included in
one of these files.
Choose the files, or drop them here
No file chosen
Choose courses.csv, terms.csv and departments_years.csv
together from the same Ally export. They are read in this
browser and stored here for the Client ID and region above,
so save those first. Nothing is uploaded to a server. Add
years.csv, months.csv and departments_terms.csv as well to
see the Trends view on your own accessibility history; all
three are needed, and the status line below says whether
history is stored.
Courses whose name carries [archived] are skipped, which is
roughly two thirds of a typical export. Search results say
when this is in effect.
For example:
https://your-institution.github.io/ally/ally-course-payload.json
The https address of the ally-course-payload.json file the
converter
produces, hosted somewhere that allows this site to read it.
A GitHub repository or GitHub Pages site does; Blackboard
does not, so a file uploaded there cannot be fetched. The
data is fetched once and stored in this browser for the
Client ID and region above, so save those first. Select the
button again to refresh it.
One ally-course-payload.json file. Data only: nothing from
the host runs on this page.
Loads ally-lookup-data.js and ally-courses-data.js from
the folder at that address. It works from any host,
including Blackboard, but it runs whoever controls that
address's code inside this site, beside every key this
site has stored for you: your Ally API token, and any
OpenRouter, MathPix or Azure keys. Choose it only for an
address your own institution controls. Nothing is stored;
the scripts are loaded again each time Ally Reporting
opens.
Chat
Type to filter the model list by name or provider. The result
count is announced.
No model selected
More filters
Prices are US dollars per million tokens. Foundry models are
hidden while a price filter or sort is on.
No models match your filters. Use Reset filters to clear them.
Model information
System prompt
Optional instructions for the model. For example: Answer in plain
English
Parameters
Balanced creativity and consistency
Nucleus sampling. At 1 it has no effect.
Controls how much the model repeats individual words. Higher
values encourage a wider vocabulary; lower values allow more
word repetition. Values below 0 encourage repeating words, which
can help keep the response consistent.
Controls how much the model repeats phrases and ideas. Higher
values encourage new phrasing and fresh topics; lower values
allow more repetition. Values below 0 encourage repetition,
which can help keep the response consistent.
Moderate length responses
Try one of these
Images (JPEG, PNG or WebP) up to 10MB, PDF up to 150MB, a text,
Markdown or LaTeX file, or a Word, Excel or PowerPoint file.
Optional, and MathPix charges for it. It reads the file and
lets you correct the text, and AI models usually do better
with the corrected text. It opens ready to upload, shows the
cost, and sends nothing until you confirm there.
These go with your message, and with each later message while
this file stays in the conversation.
Turn a recording into text. Upload an audio file and get a
transcript with timings and speaker labels.
No file chosen
Transcript
In edit mode, press Enter or move away from the box to save a
correction. Escape puts back the last saved text, discarding
anything you have typed since.
Choose a speaker, type their name and apply. Leave the box empty
and apply to put that speaker back to its number.
No lines selected
Tick the lines you want, choose who said them, then change the
speaker. Add speaker makes a new speaker for lines the
transcription did not separate.
No suggestions.
Suggestions come from an automated pass over the transcript
and are not part of what was recorded.
Suggestions that cannot be applied
Phrase
Suggested wording
Why it cannot be applied
Speaker labels are estimates. Similar voices can be merged into
one speaker.
Captions Fixer
Tidy up an automatic caption file. Upload the subtitles from a
recording and the tool suggests corrections for words the
transcription got wrong. You decide which to apply, and the timings
are never changed.
No file chosen
No file chosen
One term per line, spelled as it should appear. Leave it empty if
you have none. Terms from a slide deck land here for you to check;
remove any that are not real terms.
Suggested corrections
Each line is a word the tool thinks was misheard, and how many
captions it appears in. Nothing is ticked to start with — tick
only the corrections you agree with.
Held back by the checks
These corrections would change something the tool protects — a
number, a speaker label, or the length of a caption. You can
still tick one if you disagree.
Proposed by only one of the two runs
The tool asks twice and compares. These corrections came back
from only one of the two runs, so they are less certain than the
list above — nothing has been thrown away. You can still tick
any of them.
Changes to your captions
A change fixes a phrase that was wrong in several captions. A
suggested change comes from reading one caption on its own, so
check it against the recording.
No changes yet.
Caption
Start
Was
Now
Reason
Keep
Edit
Word forms
Rebuild a Word form so it works with a screen reader. Choose a Word
file, read what the checks found and the plan for the new form, then
agree to build it or stop.
1. Choose your form
No file chosen
On this route your form is sent through OpenRouter to one of the companies that run this model. The tool asks OpenRouter to use only services that keep nothing they are sent, but some may still keep requests to check for abuse. To rule that out, set your OpenRouter privacy setting to 'ZDR only'.
Run progress
Total
Read your form, ,
Write the plan,
Check the new form,
Fix what the check found, ,
2. What we found
3. The plan
You stopped here, so your form was not changed.
4. Your new form
What the check allowed
Before you send it
TikZ Studio
Write TikZ source and see the diagram it draws. Add a text
alternative, a caption and a long description, then download the
figure or export it. Everything runs in your browser.
TikZ source
Tab moves on to the next control. Lines that start with % @alt:,
% @caption:, % @longdesc: or % @decorative above the picture are
the text alternative, caption, long description and decorative
setting; the fields below keep them in step.