A PRACTICAL NOTE · Local models

How to apply this guide

Choose two compatible models actually offered in your edition. Use the same short audio clip for both when import is available.

Start with

Count misses in names, numbers and final words, and time processing. A smaller model may help with storage or speed but results vary.

Check before you finish

Run one noisy sample and one clear sample, both in your intended language. Correct the text and include editing time in the comparison. The useful gain is the time and storage you save after accounting for mistakes you would otherwise need to fix.

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: Choose two compatible models actually offered in your edition. Use the same short audio clip for both when import is available.

Measure or compare

Count misses in names, numbers and final words, and time processing. A smaller model may help with storage or speed but results vary.

Decide what to keep

Keep the smallest model that passes your own important phrases. Match language support before comparing size.

What to watch for

Run one noisy sample and one clear sample, both in your intended language. Correct the text and include editing time in the comparison. The useful gain is the time and storage you save after accounting for mistakes you would otherwise need to fix.

Keep exploring