In this short interview, Speaker A describes a project that placed adults with autism and Down syndrome in a warehouse setting to collect real‑world data on their speech, cognition, and physical task performance. The team used this data to train multiple AI models: a speech model that continuously adapts to each individual's atypical speech patterns, providing voice‑to‑text, text‑to‑voice, and visual feedback; and a vision/physical‑skill model that monitors workers’ movements to detect errors such as missing a stop point or failing to locate the correct bin. After achieving success, the tools were released freely for anyone aiming to improve hiring and workplace support for disabled adults. Speaker B asks for clarification that the “model” refers to artificial‑intelligence models.
We took people that had autism and Down syndrome in our warehouse and had them work there. We were training the model on how they speak, how they talk, how they listen, the way they think, their thought process, their math level, and their speed at which they walk, how much directions they need. So we just spent a lot of time watching them live, capturing the data, and retraining the model. And we had a lot of success at it. And then we just gave it away to anybody that wants to use it to help move the needle for hiring adults with disabilities.
When you say the model, you mean an artificial intelligence model? What model are we talking about?
Well, with the tools that we built, we have several models. One is a speech model. Adults with disabilities have varying degrees of speech challenges, whether they stammer, stutter, they garble when they talk, or they can't get the words out. We train the model that the more the adults with disabilities use it, the more the model understands how they speak. And we do voice-to-text and text-to-voice. So when they say something, we'll show them what the bot thought it heard, and then the bot will say it, and then the bot will send them a response. It'll say a response to them, but it'll also display what the bot is saying. So even if we got it wrong, then the child or the individual, the adult, can say it again. So we train their speech model was the first part. That was the hardest part. But once you get them to do it, you're using it over and over and over again, the bot learns how they talk. It doesn't matter what the disability, it doesn't matter how they talk, the bot will understand over time. So we trained that model. The second model we built was trying to leverage their physical skills, seeing if the bot could see the mistakes that they were making, whether they were walking past the place they were supposed to stop at, or they couldn't scan the right bin because they couldn't see it.
