How to normalize confidence values?

DeepSpeech
#1

Confidence from the metadata of the native client JSON call seems to be: “roughly the sum of the acoustic model logit values for each timestep/character that contributed to the creation of this transcription.”

So, in order to get values between 0 … 1 should I divide by the number of alphabet characters? And will this be above 90% for a “good” translation?

Any ideas, comments? What are typical good values?

Reliability of Metadata.confidence
#2

someone? would be great to have some sort of input

#3

You should only use it relatively, to compare against other confidence values. There’s no need to convert it to an absolute scale, and there’s no “good transcription” cutoff point.

1 Like
#4

Thanks, that helps. I guess the big players have their test data and “normalize” their values accordingly to 0…1