What changes for smart-glasses audio in a noisy place?
The system has to distinguish the intended voice from competing sound. Nearby speech, movement and microphone placement can change what reaches recognition. A quiet-room demonstration does not establish the same result in a busy conversation.
Use the errors to locate the problem
Note whether words are missed, another speaker is captured or the application fails to detect the end of the request. Those observations imply different problems. Treating all of them as poor AI makes it harder to improve either the audio input or the interaction design.
A useful comparison repeats the same short exchange in the intended setting at a normal speaking level. Record corrections and whether people have to take turns differently. If the workflow only succeeds when everyone changes how they speak, that adaptation is part of the product experience and should be reported.