UBM-GMM¶

Now you’re ready to run to run your speaker recognition experiment using bob. To do so, you can use the script ./bin/verify.py. One example of command line is:

$ ./bin/verify.py -d ‘ami’ -p ‘energy-2gauss’ -e ‘mfcc-60’ –algorithm ‘gmm’ -s ubm_gmm –groups {dev,eval}

In this example, the database used is amicorpus, the preprocessor is an Energy-based VAD, the feature extractor is MFCC-60 (19 MFCC coefficients + energy + first and second derivatives). UBM-GMM modeling (with 265 Gaussians).

The performance of the system on DEV and EVAL are:

• The HTER of the development set of 'UBM-GMM' is 1.877%
• The HTER of the evaluation  set of 'UBM-GMM' is 8.202%

[Min. criterion: EER] Threshold on Development set: 8.567133e-01

[Min. criterion: Min. HTER] Threshold on Development set: 8.163361e-01

JFA¶

To use JFA toolchain instead of UBM-GMM:

$ ./bin/verify.py -d ‘ami’ -p ‘energy-2gauss’ -e ‘mfcc-60’ –algorithm ‘jfa’ -s ‘jfa’ –groups {dev,eval}

The performance of the system on DEV and EVAL are:

• The HTER of the development set of 'JFA' is 1.652%
• The HTER of the evaluation set of 'JFA' is 11.561%

[Min. criterion: EER] Threshold on Development set: 7.990545e-01

[Min. criterion: Min. HTER] Threshold on Development set: 7.902645e-01

ISV¶

Another example is to use ISV toolchain, but first you need to train the ISV model using the training set (called world in bob):

$ ./bin/train_isv.py ‘ami’ -p ‘energy-2gauss’ -e ‘mfcc-60’ –algorithm ‘isv’ -s ‘isv’ –groups ‘world’ -o ‘isv’ –preprocessed-directory ‘YOUR_PREPROCESSED_DIR’ –extracted-directory ‘YOUR_EXTRACTED_FEATURES_DIR’ –gmm-directory ‘YOUR_UBM_DIR’

Then you can run the script ./bin/verify.py:

$ ./bin/verify.py -d ‘ami’ -p ‘energy-2gauss’ -e ‘mfcc-60’ –algorithm ‘isv’ -s ‘isv’ –groups {dev,eval} –skip-preprocessing –skip-extraction –skip-projector-training –preprocessed-directory ‘YOUR_PREPROCESSED_DIR’ –extracted-directory ‘YOUR_EXTRACTED_FEATURES_DIR’ –projector-file ‘DIRECTORY_OF_PROJECTOR_FILE’

The performance of the system on DEV and EVAL are:

• The HTER of the development set of 'ISV' is 1.426%
• The HTER of the evaluation set of 'ISV' is 3.360%

[Min. criterion: EER] Threshold on Development set: 2.599592e-01

[Min. criterion: Min. HTER] Threshold on Development set: 3.543569e-01

TV-cosine¶

To use Total Variability modeling (TV) or also known as i-vector using cosine distance scoring:

$ ./bin/train_ivector.py -d ‘ami’ -p ‘energy-2gauss’ -e ‘mfcc-60’ -a ‘ivector-cosine’ -s ‘ivec-cosine’ –groups ‘world’ $ ./bin/verify.py -vv -d ‘ami’ -p ‘energy-2gauss’ -e ‘mfcc-60’ -a ‘ivector-cosine’ -s ‘ivec-cosine’ –groups {dev,eval} –skip-projector-training

The performance of the system on DEV and EVAL are:

• The HTER of the development set of 'TV-cosine' is 10.586%
• The HTER of the evaluation set of 'TV-cosine' is 17.194%

[Min. criterion: EER] Threshold on Development set: 8.452344e-01

[Min. criterion: Min. HTER] Threshold on Development set: 8.646607e-01

TV-PLDA¶

To use i-vector using PLDA:

$ ./bin/train_ivector.py -d ‘ami’ -p ‘energy-2gauss’ -e ‘mfcc-60’ -a ‘ivector-plda’ -s ‘ivec-plda’ –groups ‘world’ $ ./bin/verify.py -d ‘ami’ -p ‘energy-2gauss’ -e ‘mfcc-60’ -a ‘ivector-plda’ -s ‘ivec-plda’ –groups {dev,eval} –skip-projector-training

The performance of the system on DEV and EVAL are:

• The HTER of the development set of 'TV-PLDA' is 2.252%
• The HTER of the evaluation set of 'TV-PLDA' is 4.545%

[Min. criterion: EER] Threshold on Development set: -6.309673e+01

[Min. criterion: Min. HTER] Threshold on Development set: -5.790133e+01