Voice AI

App access

Access secret

Stored only for this browser tab session. It is not written into the app code.

Session

Session label

Model

Provider

Model

Effort

Voice commands

Conversation

Recorded sessions

Every conversation session stored on this device by the recorder, newest first — including sessions from earlier page loads. Select one to read its record. Nothing here is uploaded.

Conversation record

Diagnostic timeline

Settings

Record session diagnostics

Diagnostics are stored locally on this device only. They include recognised speech and the AI's responses, alongside timing and audio measurements. Nothing is uploaded. This setting can only be changed while the session is stopped.

Conversation record

Both sides, in order, from the session recorder. Empty when recording is off.

Diagnostic timeline

Everything, ordered, generation-stamped. Refresh to render the current recording.

Calibration and test tools

Calibration

Start the session first, then pick the environment once (it is remembered). Set the phone media volume to the target and check the readout. Assistant-only = stay silent; double-talk = speak naturally over most of the ~30 s passage. Countdown, then do not touch the phone until COMPLETE. The grid tracks what this device has captured. Artifacts save through the normal Stop & Save flow.

idle

Environment (set once per location — remembered across refresh)

Media volume: —

Typed question (the assistant answers aloud; you stay silent)

Double-talk and user-only runs put STT in shadow mode: your voice is recorded but never interrupts playback, never reaches the model, and changes nothing. Press once, wait the countdown, then talk naturally.

The tap test plays the medium fixed utterance. When you deliberately knock the phone or counter, press MARK TAP at the same moment — the press is the event marker. Keep tap-test samples separate from clean baselines; they are labeled by their own mode.