A repeatable small test#
Use an invented sentence containing a name, number and negative instruction. Compare normal speech, soft speech and whispering at a normal wearing position. Repeat in a quiet room and in a permitted shared-space setting without recording private conversations.
Count important errors rather than judging only whether the text sounds fluent. Then measure editing effort and whether the corrected instruction reaches the intended conversation. Ask separately whether a nearby listener found the speech disruptive.
Add hardware checks#
Record frame comfort, connection stability, charging behavior and the specific control mapping if one is available. Note device revision, host, software version and input mode. Battery observations need a workload description; whisper observations need speaking and environmental conditions.
Do not average unrelated tests into a single “AI accuracy” score. Microphone capture, transcription, semantic refinement and task execution can each fail differently. A demonstration can establish that one route worked once; a public performance claim needs repeatable results across the conditions it promises to cover.