How to Ensure Accessibility for Streamed Content?

I'm curious about how to make streamed content accessible, especially in the context of LLMs (Large Language Models). What strategies are people using for accessibility in this new content format? Is there a possibility of introducing a new ARIA role for this?

One idea I've had is to use aria-live="polite" for individual chunks or blocks of content, so they can be queued for screen readers. However, I'm concerned because if the output takes a while, users may hear bits of information too quickly. For example, I was thinking about displaying a loading message like `

` until the whole response has streamed in.

From what I understand, setting aria-atomic="false" would cause each streamed section to be announced again, unless it's properly chunked. But I'm not exactly sure how this would work with markdown.

I have some suggestions for a potential aria-text-stream attribute:
- **whole**: Wait until the complete response is available before reading it.
- **sliced**: Break the response into segments based on a customizable character or word count.
- **bits**: Read out each word as it comes in, or use a short delay to group words together.

What do you all think? Any best practices or thoughts to share?