Confidence Calculation

[RobinE89]

Hello everyone,
I am currently working on a scientific paper on the subject of STT. It would be important for me to know how the "confidence" value is calculated inside of STT. I have looked it up in the deepspeech paper / and also in the playbook -> but unfortunately, I can't find anything precise about this and that's why I currently don´t know how to rate it (what is a good result, what is a bad one, what is the value based on...).

I initially assumed that the value describes the distance between the acoustic-model-result and the language-model-result. But I am surprised -> there is also a value when the scorer is disabled.

If I just overlooked it, I apologize(!) and thank you anyway for a helpful hint =)

best regards,
Robin