Hi Manuel,
Thanks for the information. Do you have any documentation or backgrounders on what I can do to make my results more accurate? For example I am running:
./bin/spkverif_ivector.py -d config/database/voxforge.py -p config/preprocessing/energy.py \
-f config/features/mfcc_60.py -t config/tools/ivec/ivec_256g_t100_cosine.py -z -b ivector_cosine \
--groups eval --skip-lda-train-projector --skip-lda-projection \
--skip-wccn-train-projector --skip-wccn-projection --skip-train-plda-enroler \
--skip-model-enrolment
I'm using my own data with the voxforge protocol but the rest of the settings are default. I don't really know very much about speaker ID and am learning this as I go. I'd like to know what settings I can change to perhaps make things more accurate. For example if I were to add more data to my UBM, would that help? right now I have about 500 speakers in the UBM with 8-9 seconds of speech per speaker. Would I have better accuracy if I had 2000 speakers at 8-9 seconds? or if I had 500 speakers at 30 seconds? None of the speakers in the UBM are part of the enroll or tests.
Thanks,
Maria