Recognition accuracy & confidence
How to read confidence and improve what the recognizer hears.
Improve what the recognizer hears
Accuracy has two stages: recognizing what was said, then translating it. If recognition is wrong, the translation will be wrong too — so improving recognition has the biggest payoff.
- Add recognition vocabulary for names, places, and specialized terms.
- Add translation glossary entries for idioms and phrases that translate poorly.
- Use a clean microphone and keep background noise down.
- Confirm the source language matches the speaker.
Reading the confidence indicator
The speaker dashboard shows a live confidence reading for recognition. When it drops, captions are more likely to contain errors. Treat a sustained low reading as a prompt to check your audio setup.
Low confidence usually means a noisy or distant microphone, an unusually fast speaker, or the wrong source language. Fix the input rather than the output.