How to get the result accuracy?

[njourdane]

I'm trying to code a script that detects speaker gender over time on a sound file.

I used ffmpeg to slice the file into few-seconds chunks, then I use this project to analyze each chunk.

The problem is that many of them doesn't have voice at all: it could be silences, ambiance music, various noises, etc. , on which the script still tries to guess the gender.

I tried to eliminate chunks where scores are lower than 90% but it produces many false positives (I guess it's because if the score given for one gender is bad but the score for the other one is a lot more bad, the algorithm could give a >90% certitude for the first one).

I there a way to get the accuracy of the result, or having scores for each gender independent to each other?