errors + correct do not sum to 100% - is this by design?

[chananshgong]

According to the DER tutorial here: http://www.xavieranguera.com/phdthesis/node108.html The FA + MD + CONFUSION should all add up to the DER.

However, in the example here: {u'confusion': 7.0, u'correct': 22.0, u'diarization error rate': 0.5161290322580645, u'false alarm': 7.0, u'missed detection': 2.0, u'total': 31.0} The DER is 0.5161290322580645 , and multiplying by total 31, one get 16, which indeed is the sum of confusion + false alarm + missed detection. However, adding correct (21) to errors (16) is 37 which is more than 100% (total =31)

[hbredin]

Yes. This is by design.

Also note that DER may be greater than 1 in some cases. This is due to the fact that the denominator (total) is the total duration of speech in the reference, not the total duration of the file.

[chanansh]

But why don't correct + FA + MD + confused = total?is total the reference while the sum above is of the hypothesis total?

[hbredin]

OK. Let's summarize:

• false alarm = reference non-speech regions tagged as speech in hypothesis
• missed detection = reference speech regions tagged as non-speech in hypothesis
• confusion = reference and hypothesis agree that this is a speech region, but hypothesis is wrong on the identity of the speaker
• correct = reference and hypothesis agree that this is a speech region and hypothesis is right about the identity of the speaker

Overall, you can sum those values as follows:

• correct + false alarm + confusion = total speech duration in hypothesis
• correct + missed detection + confusion = total speech duration in reference

[chanansh]

Thanks. Can you please add this to the documentation or comment in the code? Thanks for contributing this package.

[hbredin]

I'd be happy to merge a pull request updating DiarizationErrorRate's docstring.