Meeting minutes
Accessibility of Machine Learning and Generative AI.
Janina: no current new news. It's quietly bubbling along.
Machine learning and generative AI: observations and conclusions.
Jasonjgw: didn't exhaust this topic last time. Opening up this topic from last time. Any additional observations?
Janina: wondering about redirects. Haven't found ability to find redirect. Maybe better on desktop insetead of phone? Do this job, send me the output in an email, put in HTML...have we seen that anywhere?
jasonjgw: I think - and maybe time to think and discuss - tools you can install locally, including from a shell prompt, that access the LLM by running locally or remote (well known propiretary providers) and access over an API. Those tools have greater permission to edit, etc you might have in mind.
Janina: or power of command line to redirect
jasonjgw: yes. Integrations with various text editors with LLM to get access to the file and directory that the file is stored in, and other files present. It can gain a larger context, looking at work environment and files in repository
Janina: Think we may critize the industry for not doing the basics, yet
… the UI is what made it a success. It wouldn't be succesful for individula users if not for the GUI, most users won't work command line
jasonjgw: it has to be gov, edu users in large scale...bringing in revenue. That's where you get more serious integrations with local OS and tools user is running because used for work purposes, not a consumer product at that point. Tells me for our doc and work, we need to know more about some of the client side tools that aren't browser based but
using web technology using an API call
Janina: yes. interested in expanding to API. Big future, esp for AI, tie into agentic tools for particular needs that meet accessibiltiy constraints. use APIs to minimize hallucination
jasonjgw: if time, should look at tools available
Janina: on my to do list
jasonjgw: had heard you can put money in and pay as you go, costs cents...Not exactly expensive
Janina: share time model
jasonjgw: when i was undergrad - somebody at the uni was paying - but have an account on central unit they charged by CPU time and disc space you were using. Wasn't enough for anyone to complain.
… so we're back there (with the model). Reading an article summary this morning referring to LLM based tool that could be downloaded. Something to follow up
Janina: anybody with any experience in non-text AI? Anyone done anything with images or audio.
jasonjgw: I've done some text to speech ones. Some models will perform text to speech conversion process. Tried some online with sample websites to enter text and get an audio file of output. Performance is a problem, especially if you want to use it interactively. Convert a book or doc to a high quality spoke audio file then these would be ideal.
Trying to use for screen reader or interactive system where responsiveness is important, it's not there yet
Janina: i have an NYT subscrtiption, if i want to do more than read headlines, I tend to hit the listen to this article button, which gives discalimer that read by an autmated voice. But the aut voice is pretty good, but doesn't get "ST" correct
jasonjgw: also heard experiments with LLM to generate textile graphics, and have a display capable, it will display there. I haven't played with this personally.
Janina: find out who is working on this
jasonjgw: dot working on this? AI based textile graphics tool. Not sure what it's capable of doing. Should note, discussing Monarch Braille device, it's a 96 x40 pin array on display and represent textile graphics as well as braille, can load a specially prepared image file and display in textile form including braille labels. Zoom in and out of
textile graphics, apend on the display, various operations that can be perfromed
jpaton: in terms of syntehtic speech, peoples reaction to syntehtic pseech to auidio description. Took 6 clips, created audio description and tried to get people to rate by nautralness without telling them it's synthetic. Wanted to see how much it was affecting things. Many rated badly for emotional relevance to the content, but weren't picking up
that it was syntehtic speech.
Jpaton: in UK, guidelines for audio descriiption have a neutral and blank expression. People wanted emotive and humanness. Which isn't the sort of thing syntehtic speech can easily provide. Shared the link in the IRC (above)
jasonjgw: I heard some comments (podcast?) suggesting that some of the speech recognition tools based on LLMs were making different kinds of errors than expected, leading to plausible text, never actually said in a meeting, it can be misleading to that extent. It will onfabulate what it thinks someone said and might look plausible but not an
accutate transcription.
Janina: none of us have really gotten away from text. But still seem to be text based AI. And that's fine. Are we all aware of upcoming W3C workshop on voice.
Janina: to find the link. It's in Feb. Too late to submit papers.
jasonjgw: old approach where a script and a program is dead
… (I think)
jpaton: it's all probablity based. When going through system, it's most probable based on the data it has. That's where the hallucination is coming from
jasonjgw: if it starts generating false text, people who are depending on captions or transcripts are going to be seriously disadvantaged
<janina> https://
jasonjgw: Anything else regarding machine learning experiences? I think this topic keeps giving. We can bring back next week.
Miscellaneous topics.
jasonjgw: I can say, not related but close. Canadian standards group working on trust meter work that I'm involved in, making progress. Becoming very interesting. Working towards documentation and when there's a draft to share, will share with this group.
<janina> W3C Breakouts Day 2026 - CfP <https://
Janina: there's a W3C - looking at a virtual confab in March. WOrkshop proposals are solicited link is shared here
Janina: CSUN is coming up
stacey: axecon coming up
Janina: will be presenting with Paciello
Jasonjgw: no more on CAPTCHA right now?
Janina: no