Interactive Demo

Watch it listen.

A reconstructed session played in real time. WordBridge listens passively — detecting word-finding difficulty, surfacing ranked candidates and contextual anchors, and governing intervention from physiological signals. Reconstructed for illustration; no real patient audio.

READY
Anomic aphasia · morning routine0:00 / 0:19
Conversation
◖ Audio · Top-3—
Listening for word-finding difficulty…
⊹ Contextual anchor
"Talking with David, your husband — about breakfast."
74bpm
HEART RATE
GSR18%
TIER 1 · BASELINE
Press play to begin the session
Interaction Model

Processes audio in 200ms micro-turns — detecting hesitation and circumlocution in real time, without waiting for the sentence to end.

How the dual-model split works →
Background Model

Holds the full session context. Generates contextual anchors and reasons over whether physiological signals reflect genuine distress.

What makes TML-Interaction-Small different →
Safety Tier System
T1Full operationWord candidates + contextual anchors
T2Mild distressWord candidates suppressed — anchors only
T3High distressAll output suspended — silent caregiver alert

Transitions are decided by the Background Model — not raw HR/GSR thresholds alone.

Reconstructed session. The conversation script, character names, timing, and physiological values are simulated to illustrate how WordBridge would behave under each condition. No real patient data, clinical audio, or wearable sensor output was used. For a full description of the hypotheses and evaluation methodology, see the series intro.