Community · E2 · artifact verified

Find the word aphasia can't say

For people with aphasia who know what they mean but can't get the word out: describe it any way you can, and Jev picks the best guesses from a fixed 900-word list as big tap-to-hear picture tiles - never inventing a word.

01 · Role in the system

What Jev does here

After a stroke, many people with aphasia can describe the hot thing you drink in the morning but not say coffee. The browser's built-in speech recognition transcribes them talking around the word, and Jev answers multiple-choice questions over it - the vocabulary is split into six slices so one request asks all of them at once, and a second request picks among the roughly two dozen best candidates - returning guesses in about half a second to a second. The trust rules are structural: Jev can only choose from a fixed list of about 900 everyday words plus the person's own names, pets, and places; nothing is spoken until the person taps a tile; and when the description is too vague it answers not-sure-yet-keep-going instead of guessing. Tiles use ARASAAC speech-therapy symbols with browser text-to-speech, and personal words are stored only in that browser.

Choosing from 900 words is the safety property

The design constraint does the safety work: Jev can only choose from a fixed list of about 900 everyday words plus the person's own names, pets, and places — never inventing, never free-forming. A person with aphasia describes the word they cannot reach, the browser's speech recognition transcribes the description, and Jev answers multiple-choice questions over it — the vocabulary split into six slices asked in one request, then a runoff among the best two dozen. Guesses arrive in half a second to a second as big tap-to-hear picture tiles.

The confirmation loop is human by construction: nothing is spoken until the person taps a tile, and when the description is too vague the system answers not-sure-yet-keep-going instead of forcing a guess. For an assistive tool, declining to guess is the feature.

A demo that names its own distance from a device

The README's first warning is quoted in the record: this is a demo, not a medical device, and it has not yet been tested with people who have aphasia — the population it is built for. The closed vocabulary bounds what can be found by design; out-of-vocabulary needs stay out of reach, and personal words live only in that browser.

As the index's only accessibility record, it demonstrates the closed-vocabulary pattern carrying human stakes: the same enumerate-then-choose shape as a game harness, applied to speech therapy, with the honesty to say the trial with real users has not happened yet.

02 · Control boundary

Where Jev sits

Speech-to-text to a fixed closed vocabulary split into choice slices; one batched multi-slice request plus a runoff choice; the person always confirms by tapping before anything is spoken.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

  • The README's own warning, stated first: a demo, not a medical device, and it has not yet been tested with people who have aphasia.
  • The closed 900-word vocabulary plus personal words bounds what can be found; out-of-vocabulary needs stay out of reach by design.
  • Guess quality depends on the description given; vague descriptions return not-sure rather than forced guesses.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.