Test timing once#
Speak the same sentence after capture is visibly active. Compare it with an attempt where you spoke immediately. Keep device and speaking volume constant. A practical pause can help avoid talking before capture begins, but it is not a verified fix for every missing-word case.
If the beginning is still absent, check the input device and local model state. Ordinary speech should work before you test whispering. Note whether the app was newly launched or the model was still preparing.
Preserve the intended instruction#
A missing opening phrase can change scope. “Do not publish this” losing its first words becomes dangerous as an instruction. Repair the draft before delivery rather than letting refinement guess the missing text.
The reviewed materials do not establish a universal startup-latency number. A useful report includes the trigger, input, software version and the shortest sentence showing the omission. Avoid treating competitor startup behavior as a CueSpec diagnosis.