Compare the stages#
Check the raw draft for names, numbers, units and negation. If those are already wrong, start with input quality and recognition. If the raw draft is right but a refined version is wrong, evaluate refinement separately.
Repeat a short sentence at normal volume with the same input device. Then change only one condition, such as microphone position or room noise. For exact identifiers, type or paste the intended string rather than repeatedly saying it louder.
Keep a useful example#
Record the expected text, actual text and setup, using invented content where possible. Include whether the error appears consistently or only in a particular environment. A single error does not establish a general language limit, while one clean sentence does not establish full accuracy.
CueSpec’s editable draft provides a place to repair the request. It is not a guarantee that every voice will be recognized at whisper volume. If repair takes longer than typing for this task, use typed input and retain the specific issue for later investigation.