A PRACTICAL NOTE · Local models

How to apply this guide

Make a short test clip with two seconds of room sound after one spoken sentence. Compare the result with a clip of room sound only.

Start with

Write down unexpected words in both cases. A model that turns room noise into speech can produce plausible text that was never said.

Check before you finish

When comparing models, keep the length of the silence the same in both samples. One model may produce a different pattern at the end of an otherwise clear recording. Review the audio around any unexpected sentence before treating the generated text as a quotation.

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: Make a short test clip with two seconds of room sound after one spoken sentence. Compare the result with a clip of room sound only.

Measure or compare

Write down unexpected words in both cases. A model that turns room noise into speech can produce plausible text that was never said.

Decide what to keep

Discard unsupported text. Reduce noise and record shorter clips before comparing a different model.

What to watch for

When comparing models, keep the length of the silence the same in both samples. One model may produce a different pattern at the end of an otherwise clear recording. Review the audio around any unexpected sentence before treating the generated text as a quotation.

Keep exploring