Streaming announcement

Streamed text breaks the assumption a live region is built on: that an update is an event. A generated reply is not one update, it is two hundred, and a region wired the obvious way announces every one of them. The reader hears the first two words of a sentence, then the first three, then a fragment of the fourth as the queue jumps ahead, and by the time the answer is on screen the speech has said nothing usable at all. That failure mode has a name in practice, the token flood, and it is the single most common accessibility bug in a natural language interface.

The fix is to separate the writing from the announcing. Let the tokens land in the transcript as they arrive, because a sighted reader and a braille reader following along both want that, and keep the region that speaks quiet until the answer settles. This is one of the few places busy state genuinely earns its keep: aria-busy="true" on the region for the duration of the generation, cleared when the last token lands, is exactly the “this is halfway written, do not read it yet” signal streaming needs, and it is one attribute rather than a custom queue. What speaks at the end is one status message: the reply is complete, and usually how long it is, so a reader knows whether to read on or move past.

The choice you actually have to make is what “complete” means. Waiting for the whole answer is right for a short reply and wrong for a long one, because a reader who asked a question two minutes ago has been staring at silence. Announcing at sentence or paragraph boundaries is the middle ground, and it holds up as long as the boundary is real: chunking every N tokens produces the same flood, just slower. Whatever you pick, say when generation starts as well (“generating a reply”), because the gap between the question and the answer is otherwise indistinguishable from nothing having happened.

Two smaller things worth wiring while you are there. Give the reader a way to stop generation and announce that it stopped, since a stream that ends early is a state, not an absence. And do not put the streaming text itself inside an aria-live container “just to be safe”: a region that is both the transcript and the announcer is the flood by construction, and no amount of politeness setting will save it.