Added support voicePA and moved extractors to bob.bio.spear

bob/pad/voice/config/algorithm/gmm_tomi.py

-#!/usr/bin/env python
-
-import bob.bio.gmm
-
-algorithm = bob.bio.gmm.algorithm.GMMRegular(
-    number_of_gaussians = 512,
-    kmeans_training_iterations = 10,   # Maximum number of iterations for K-Means
-    gmm_training_iterations = 10,      # Maximum number of iterations for ML GMM Training
-
-)

bob/pad/voice/config/algorithm/gmm_tomi_scfc.py

-#!/usr/bin/env python
-
-import bob.bio.gmm
-
-algorithm = bob.bio.gmm.algorithm.GMMRegular(
-    number_of_gaussians = 512,
-    kmeans_training_iterations = 10,   # Maximum number of iterations for K-Means
-    gmm_training_iterations = 10,      # Maximum number of iterations for ML GMM Training
-    training_threshold = 5e-5,         # Threshold to end the ML training, make it smaller
-    variance_threshold = 5e-7,         # Minimum value that a variance can reach, make it smaller than default
-
-)

bob/pad/voice/config/algorithm/isv_avspoof.py

-import bob.bio.gmm
-
-algorithm = bob.bio.gmm.algorithm.ISV(
-    # ISV parameters
-    subspace_dimension_of_u = 50,
-    # GMM parameters
-    number_of_gaussians = 256,
-    training_threshold = 0.0,
-)

bob/pad/voice/config/algorithm/ivec_avspoof.py

-import bob.bio.gmm
-
-algorithm = bob.bio.gmm.algorithm.IVector(
-    # IVector parameters
-    subspace_dimension_of_t = 100,
-    update_sigma = True,
-    tv_training_iterations = 25,  # Number of EM iterations for the TV training
-    # GMM parameters
-    number_of_gaussians = 256,
-    training_threshold = 0.0,
-    use_lda = True,
-    use_wccn = True,
-    use_plda = True,
-    lda_dim = 50,
-    plda_dim_F  = 50,
-    plda_dim_G = 50,
-    plda_training_iterations = 200,
-)

bob/pad/voice/config/database/asvspoof.py

@@ -4,8 +4,7 @@ import bob.pad.voice.database
 
 
 # directory where the wave files are stored
-asvspoof_input_dir = "/idiap/resource/database/ASVspoof2015/ASVspoof2015_development"  #Path to ASVspoof in Idiap
-# asvspoof_input_dir = "/Users/pavelkor/Documents/pav/idiap/data/asvspoof"  #path to a local copy
+asvspoof_input_dir = "[YOUR_AVSPOOF_WAV_DIRECTORY]"
 asvspoof_input_ext = ".wav"
 
 

bob/pad/voice/config/database/asvspoof_verify.py

-#!/usr/bin/env python
-
-import bob.bio.base.database
-
-
-# directory where the wave files are stored
-# asvspoof_input_dir = "/idiap/resource/database/mobio/denoisedAUDIO_16k/"  #MOBIO database for UBM training
-asvspoof_input_dir = "/idiap/resource/database/ASVspoof2015/ASVspoof2015_development"  #Path to ASVspoof in Idiap
-# asvspoof_input_dir = "/Users/pavelkor/Documents/pav/idiap/data/asvspoof"  #path to a local copy
-asvspoof_input_ext = ".wav"
-
-
-database_licit = bob.bio.base.database.ASVspoofBioDatabase(
-    protocol='licit',
-    original_directory=asvspoof_input_dir,
-    original_extension=asvspoof_input_ext,
-    training_depends_on_protocol=True,
-)
-
-database_spoof = bob.bio.base.database.ASVspoofBioDatabase(
-    protocol='spoof',
-    original_directory=asvspoof_input_dir,
-    original_extension=asvspoof_input_ext,
-    training_depends_on_protocol=True,
-)

bob/pad/voice/config/database/avspoof.py

@@ -4,8 +4,7 @@ import bob.pad.voice.database
 
 
 # directory where the wave files are stored
-avspoof_input_dir = "/idiap/project/lobi/AVSpoof/data/"  # folder at Idiap
-# avspoof_input_dir = "/Users/pavelkor/Documents/pav/idiap/data/AVspoof_full/data/"  # local copy
+avspoof_input_dir = "[YOUR_AVSPOOF_WAV_DIRECTORY]"
 avspoof_input_ext = ".wav"
 
 

bob/pad/voice/config/database/avspoof_verify.py

-
-#!/usr/bin/env python
-
-import bob.bio.base.database
-
-# directory where the wave files are stored
-avspoof_input_dir = "/idiap/project/lobi/AVSpoof/data/"
-# avspoof_input_dir = "/Users/pavelkor/Documents/pav/idiap/data/AVspoof_full/data/"
-avspoof_input_ext = ".wav"
-
-
-database_licit = bob.bio.base.database.AVspoofBioDatabase(
-    protocol='licit',
-    original_directory=avspoof_input_dir,
-    original_extension=avspoof_input_ext,
-    training_depends_on_protocol=True,
-)
-
-database_spoof = bob.bio.base.database.AVspoofBioDatabase(
-    protocol='spoof',
-    original_directory=avspoof_input_dir,
-    original_extension=avspoof_input_ext,
-    training_depends_on_protocol=True,
-)

bob/pad/voice/config/database/voicepa.py

+#!/usr/bin/env python
+
+import bob.pad.voice.database
+
+
+# directory where the wave files are stored
+voicepa_input_dir = "[YOUR_VOICEPA_WAV_DIRECTORY]"
+voicepa_input_ext = ".wav"
+
+
+database = bob.pad.voice.database.VoicePAPadDatabase(
+    protocol='greattest',
+    original_directory=voicepa_input_dir,
+    original_extension=voicepa_input_ext,
+    training_depends_on_protocol=True,
+)

bob/pad/voice/config/extractor/imfcc20.py

-
-import bob.pad.voice.extractor
-
-extractor = bob.pad.voice.extractor.CepstralExtended(
-    # the parameters emulate the ones in the paper "A Comparison of Features for Synthetic Speech Detection" by
-    # Md Sahidullah, Tomi Kinnunen, Cemal Hanilci
-
-    pre_emphasis_coef=0.97, # as per the algorithm implemented in the paper
-    n_ceps=20, # number of Inversed MFCC coefficients should be 20 - this is the number we cut off from whole Cepstrum
-    n_filters=20, # number of filters in the bank is also 20
-    win_length_ms=20.,  # 20 ms - this is the value in the paper
-    win_shift_ms=10.,  # This is the overlap - half of the window
-    f_max=8000,  # this number insures we take half of the frequencies after FFT - so we retain only 257 values for 512 window
-    mel_scale=True,  # Mel-scaling is what make these IMMFC features
-    with_delta=True,  # As reported in the paper
-    with_delta_delta=True,  # As reported in the paper
-    energy_filter =True, # The paper uses power of FFT magnitude
-    dct_norm = True, # The paper uses normed DCT-II variant
-    delta_win=1, # the paper computes deltas on window of size 1
-    inverse_filter=True, # this uses inversed Mel-filter
-)
-

bob/pad/voice/config/extractor/lbp_eurecom.py

 
-import bob.pad.voice.extractor
+import bob.bio.spear.extractor
 
 """
 This LBP-based feature extraction is implemented based on the paper
@@ -13,7 +13,7 @@ of the signal. Histogram size is 58 (as lbp_uniform LBP8,1 has 58 different valu
 GLCMs are not computed for top and bottom row of the spectrogram. Hence, the resulted size of the feature vector is 2842.
 """
 
-lfcc16_eurecom = bob.pad.voice.extractor.CepstralExtended(
+lfcc16_eurecom = bob.bio.spear.extractor.CepstralExtended(
     # the parameters are as specified in the paper by Eurecom
     # Federico Alegre, Ravichander Vipperla, Asmaa Amehraye, Nicholas Evans.
     # "A new speaker verification spoofing countermeasure based on local binary patterns".

bob/pad/voice/config/extractor/lfcc20.py

-
-import bob.pad.voice.extractor
-
-extractor = bob.pad.voice.extractor.CepstralExtended(
-    # the parameters are as specified in the paper "A Comparison of Features for Synthetic Speech Detection" by
-    # Md Sahidullah, Tomi Kinnunen, Cemal Hanilci
-
-    pre_emphasis_coef=0.97, # as per the algorithm implemented in the paper
-    n_ceps=20, # number of LFCC coefficients should be 20 - this is the number we cut off from whole Cepstrum
-    n_filters=20, # number of filters in the bank is also 20
-    win_length_ms=20.,  # 20 ms - this is the value in the paper
-    win_shift_ms=10.,  # This is the overlap - half of the window
-    f_max=8000,  # this number insures we take half of the frequencies after FFT - so we retain only 257 values for 512 window
-    mel_scale=False,
-    with_delta=True,  # As reported in the paper
-    with_delta_delta=True,  # As reported in the paper
-    energy_filter =True, # The paper uses power of FFT magnitude
-    dct_norm = True, # The paper uses normed DCT-II variant
-    delta_win=1, # the paper computes deltas on window of size 1
-)
-

bob/pad/voice/config/extractor/mfcc20.py

-
-import bob.pad.voice.extractor
-
-extractor = bob.pad.voice.extractor.CepstralExtended(
-    # the parameters emulate the ones in the paper "A Comparison of Features for Synthetic Speech Detection" by
-    # Md Sahidullah, Tomi Kinnunen, Cemal Hanilci
-
-    pre_emphasis_coef=0.97, # as per the algorithm implemented in the paper
-    n_ceps=20, # number of MFCC coefficients should be 20 - this is the number we cut off from whole Cepstrum
-    n_filters=20, # number of filters in the bank is also 20
-    win_length_ms=20.,  # 20 ms - this is the value in the paper
-    win_shift_ms=10.,  # This is the overlap - half of the window
-    f_max=8000,  # this number insures we take half of the frequencies after FFT - so we retain only 257 values for 512 window
-    mel_scale=True,  # Mel-scaling is what make these MMFC features
-    with_delta=True,  # As reported in the paper
-    with_delta_delta=True,  # As reported in the paper
-    energy_filter =True, # The paper uses power of FFT magnitude
-    dct_norm = True, # The paper uses normed DCT-II variant
-    delta_win=1, # the paper computes deltas on window of size 1
-)
-

bob/pad/voice/config/extractor/rfcc20.py

-
-import bob.pad.voice.extractor
-
-extractor = bob.pad.voice.extractor.CepstralExtended(
-    # the parameters emulate the ones in the paper "A Comparison of Features for Synthetic Speech Detection" by
-    # Md Sahidullah, Tomi Kinnunen, Cemal Hanilci
-
-    pre_emphasis_coef=0.97, # as per the algorithm implemented in the paper
-    n_ceps=20, # number of RFCC coefficients should be 20 - this is the number we cut off from whole Cepstrum
-    n_filters=20, # number of filters in the bank is also 20
-    # win_length_ms=32.,  # 32 ms - this insures the window size is 512 and also FFT size is 512
-    # win_shift_ms=16.,  # This is the overlap - half of the window
-    win_length_ms=20.,  # 20 ms - this is the value in the paper
-    win_shift_ms=10.,  # This is the overlap - half of the window
-    f_max=8000,  # this number insures we take half of the frequencies after FFT - so we retain only 257 values for 512 window
-    mel_scale=False,
-    rect_filter=True, # these features are computed with a rectangular filter, hence the name RFCC
-    with_delta=True,  # As reported in the paper
-    with_delta_delta=True,  # As reported in the paper
-    energy_filter =True, # The paper uses power of FFT magnitude
-    dct_norm = True, # The paper uses normed DCT-II variant
-    delta_win=1, # the paper computes deltas on window of size 1
-)
-

bob/pad/voice/config/extractor/scfc20.py

-
-import bob.pad.voice.extractor
-
-extractor = bob.pad.voice.extractor.CepstralExtended(
-    # the parameters are as specified in the paper "A Comparison of Features for Synthetic Speech Detection" by
-    # Md Sahidullah, Tomi Kinnunen, Cemal Hanilci
-    # SCFC features as per paper that compares different features
-
-    pre_emphasis_coef=0.97, # as per the algorithm implemented in the paper
-    n_ceps=20, # number of SCFC coefficients should be 20 - this is the number we cut off from whole Cepstrum
-    n_filters=20, # number of filters in the bank is also 20
-    win_length_ms=20.,  # 20 ms - this is the value in the paper
-    win_shift_ms=10.,  # This is the overlap - half of the window
-    f_max=8000,  # this number insures we take half of the frequencies after FFT - so we retain only 257 values for 512 window
-    mel_scale=False,
-    with_delta=True,  # As reported in the paper
-    with_delta_delta=True,  # As reported in the paper
-    rect_filter=True, # in the paper, these features are computed with a rectangular filter
-    scfc_features=True,
-    log_filter=False, # it is important to set it to False, so no log is applied to SCFC features
-    # energy_filter =True, # The paper DOES NOT use power of FFT magnitude for SCMC and SCFC features
-    delta_win=1, # the paper computes deltas on window of size 1
-)
-

bob/pad/voice/config/extractor/scmc20.py

-
-import bob.pad.voice.extractor
-
-extractor = bob.pad.voice.extractor.CepstralExtended(
-    # the parameters are as specified in the paper "A Comparison of Features for Synthetic Speech Detection" by
-    # Md Sahidullah, Tomi Kinnunen, Cemal Hanilci
-    # SCMC features as per paper that compares different features
-
-    pre_emphasis_coef=0.97, # as per the algorithm implemented in the paper
-    n_ceps=20, # number of SCMC coefficients should be 20 - this is the number we cut off from whole Cepstrum
-    n_filters=20, # number of filters in the bank is also 20
-    win_length_ms=20.,  # 20 ms - this is the value in the paper
-    win_shift_ms=10.,  # This is the overlap - half of the window
-    f_max=8000,  # this number insures we take half of the frequencies after FFT - so we retain only 257 values for 512 window
-    mel_scale=False,
-    with_delta=True,  # As reported in the paper
-    with_delta_delta=True,  # As reported in the paper
-    rect_filter=True, # in the paper, these features are computed with a rectangular filter
-    scmc_features=True,
-    # energy_filter =True, # The paper DOES NOT use power of FFT magnitude for SCMC and SCFC features
-    dct_norm = True, # The paper uses normed DCT-II variant
-    delta_win=1, # the paper computes deltas on window of size 1
-)
-

bob/pad/voice/config/extractor/ssfc20.py

-
-import bob.pad.voice.extractor
-
-extractor = bob.pad.voice.extractor.CepstralExtended(
-    # SSFC features as per paper that compares different features
-
-    pre_emphasis_coef=0.97, # as per the algorithm implemented in the paper
-    n_ceps=20, # number of SSFC coefficients should be 20 - this is the number we cut off from whole Cepstrum
-    n_filters=20, # number of filters in the bank is also 20
-    win_length_ms=20.,  # 20 ms - this is the value in the paper
-    win_shift_ms=10.,  # This is the overlap - half of the window
-    f_max=8000,  # this number insures we take half of the frequencies after FFT - so we retain only 257 values for 512 window
-    mel_scale=False,
-    with_delta=True,  # As reported in the paper
-    with_delta_delta=True,  # As reported in the paper
-    rect_filter=True, # in the paper, these features are computed with a rectangular filter
-    ssfc_features=True,
-    normalize_spectrum=True, # normalized spectrum is used to compute SSFC features
-    energy_filter=False, # The Spectral Flux does not take the power of FFT itself but of the difference
-    dct_norm = True, # The paper uses normed DCT-II variant
-    delta_win=1, # the paper computes deltas on window of size 1
-)
-

bob/pad/voice/config/grid/modest.py

-import bob.bio.base
-
-# define a queue with modest parameters
-grid = bob.bio.base.grid.Grid(
-  number_of_preprocessing_jobs=48,
-  number_of_extraction_jobs=48,
-  number_of_projection_jobs=48,
-  number_of_enrollment_jobs=48,
-  number_of_scoring_jobs=1,
-  training_queue = '32G',
-  # preprocessing
-  preprocessing_queue = '4G-io-big',
-  # feature extraction
-  extraction_queue = '4G-io-big',
-  # feature projection
-  projection_queue = '4G-io-big',
-  # model enrollment
-  enrollment_queue = '8G',
-  # scoring
-  scoring_queue = '8G-io-big'
-)

bob/pad/voice/database/__init__.py

@@ -3,6 +3,7 @@ from .asvspoof import ASVspoofPadDatabase
 from .avspoof import AVspoofPadDatabase
 from .replay import ReplayPadDatabase
 from .replaymobile import ReplayMobilePadDatabase
+from .voicepa import VoicePAPadDatabase
 
 # gets sphinx autodoc done right - don't remove it
 __all__ = [_ for _ in dir() if not _.startswith('_')]

bob/pad/voice/database/asvspoof.py

@@ -4,9 +4,8 @@
 # Tue 17 May 15:43:22 CEST 2016
 
 """
-  ASVspoof database implementation of bob.db.verification.utils.Database interface.
-  It is an extension of an SQL-based database interface, which directly talks to ASVspoof database, for
-  verification experiments (good to use in bob.pad.base framework).
+  This is a high level interface for presentation attack ASVspoof database.
+  It is an extension of an interface defined inside bob.pad.base PAD framework.
 """
 
 from bob.pad.voice.database import PadVoiceFile