How percentages are calculated

Calculation

Percent false positives and false negatives are defined as the proportion of samples the model failed on within the half of the data that was marked as either positive or negative. So:

• % false positives: false_positives / total_negatives where total_negatives = false_positives + true_negatives
• % false negatives: false_negatives / total_positives where total_positives = false_negatives + true_positives

For reference, here's the definitions of all these terms:

• false positive: Was not a wake word, but model incorrectly predicts it was
• true negative: Was not a wake word, and model correctly predicts it was not
• false negative: Was a wake word, but model incorrectly predicts it was not
• true positive: Was a wake word, and model correctly predicts it was

Example

Given the following output:

=== Counts ===
False Positives: 1
True Negatives: 819
False Negatives: 6
True Positives: 43

=== Summary ===
862 out of 869
99.19 %

0.12 % false positives
12.24 % false negatives

The percentages were calculated as follows:

• 1 / (1 + 819) = 0.12% and
• 6 / (6 + 43) = 12.24%