livesay demo github

accessibility × streaming

Your server counts tokens. livesay counts ears.

A screen reader turns DOM updates into speech. Stream an answer token by token and the listener hears a stutter that never finishes. livesay records what the page announces, simulates what a screen reader would say, and tells you in numbers — announcements, partials, duplicates, silence — before your users do.

npx playwright install chromium
npx livesay audit http://127.0.0.1:4319/bad.html --html report.html

Naive streaming

What most chat UIs ship: the answer body is a live region.

livesay audit .../bad.html
Open the naive demo →

With the announcer

The same stream: no live region on the body, a completion signal, one assertive error.

livesay audit .../good.html --budget "announcements<8"
Open the corrected demo →
Runs each demo with the probe and counts what a listener would hear (nvda-like profile).
naive—
livesay—