PYLLR

Python toolkit for likelihood-ratio calibration of binary classifiers

The emphasis is on binary classifiers (for example speaker verification), where the output of the classifier is in the form of a well-calibrated log-likelihood-ratio (LLR). The tools include:

• PAV and ROCCH score analysis.
• DET curves and EER
• DCF and minDCF
• Bayes error-rate plots
• Cllr
• Fusion and calibration are not presently implemented, but we will do so soon.

Most of the algorithms in PYLLR are Python translations of the older MATLAB BOSARIS Tookit. Descriptions of the algorithms are available in:

Niko Brümmer and Edward de Villiers, The BOSARIS Toolkit: Theory, Algorithms and Code for Surviving the New DCF, 2013.

Out of a hundred trials, how many errors does your speaker verifier make?

We have incluced in the examples directory, some code that reproduces the plots in our paper:

Niko Brümmer, Luciana Ferrer and Albert Swart, "Out of a hundred trials, how many errors does your speaker verifier make?", 2011, https://arxiv.org/abs/2104.00732.

For instructions, go to the readme.

Install

Install using pip, directly from this github repository:

pip install git+https://github.com/bsxfan/PYLLR.git