Plot Kaldi confidences against other measures

[fako]

Such as:

• [ ] source
• [ ] duration
• [ ] language
• [x] speaker segments
• [ ] silences
• [ ] WER
• [ ] music segment count

[fako]

Speaker segment probability plots can be found here: https://github.com/SURFpol/pol-harvester/blob/master/research/kaldi_confidence/Segment%20confidences.ipynb It's striking that all probabilities are low. This does not predict much good for making decisions about for instance language based on the probabilities.