A PRACTICAL NOTE · Local models

How to apply this guide

Run the same short sentence just after launching, then once more after the model has already processed a clip.

Start with

Measure the time from Finish to the first complete result for each run. Record whether model loading or a download was visible.

Check before you finish

Take three timings after the first result and compare their middle value with the first-run time. If the device was idle for several minutes in your normal workflow, add a test after that gap. The same model can take different time depending on whether it remains loaded.

Evidence to keep: Write down the speech model, language and app edition; change one of them at a time when comparing results. This page’s key limit is: Run the same short sentence just after launching, then once more after the model has already processed a clip.

Measure or compare

Measure the time from Finish to the first complete result for each run. Record whether model loading or a download was visible.

Decide what to keep

Judge repeated delay separately from the first launch. Keep the first-run wait in mind for occasional use.

What to watch for

Take three timings after the first result and compare their middle value with the first-run time. If the device was idle for several minutes in your normal workflow, add a test after that gap. The same model can take different time depending on whether it remains loaded.

Keep exploring