Live session
00:00Audio goes to model providers. Sessions stop after 15 minutes. Design actions are simulated.
Decisions
Start listening to test a variant.
Replay lab
Transcript comparisons isolate judgment; live audio replay also tests transcription and timing. Labels are editable below.
Transcript & labels
Method & limits
Generated scenes are development fixtures with authored labels, not a held-out accuracy test. Related voices and scripts are not independent samples. Synthetic comparisons use each variant’s own earlier decisions. Recorded comparisons reuse captured past context. Timing variants require audio replay. API failures are reported separately. Audio replay uses 20 ms frames at normal speed. Uploaded and recorded audio stays on this device until replayed to providers; export it before clearing browser storage. Compare “Audio + text” against “Gemini text control” to isolate audio. Initial context is supplied as text, never played into the microphone stream.