It's not possible to use verify_gmm with the algorithm GMMRegular (the one that uses the traditional scoring ).

It seems that such algorithm was not designed to used the distributed implementation. If you set it, the algorithm will try to train the UBM twice, one distributed and one not distributed (during the train-enroller; and this one will fail).

The code below with the attached file will raise the issue.

verify_gmm.py -r 8 --run-local-scheduler vox_forge_test.py

Thanks Nathalia for reporting it (She's not part of gitlab so I can't use the @, bit she can see the issue at least).

vox_forge_test.py