def do_fishers_pretrained_ubm():
feature_dir = work_dir + 'data/{}/'.format(recipe)
out_dir = work_dir + 'data/' # Where the computed features will be saved
ubm_dir = work_dir + 'data/UBMs/' + ubm_folder_name + '/ivec_models/' # where the diagonal ubms live
# list_ubm_files = util.traverse_dir(ubm_dir, '.dubm') # reading all the files with .mdl or .dubm as format (latter is more reliable)
list_sets = ['dementia_new8k']
for g in list_n_clusters:
for deltas in [0, 1, 2]:
# info-purpose parameters from the frame-level extracted features #
feats_info = [
20, deltas, 'mfcc'
] # info of the features (n_features/dimension, deltas, cepstral_type=choose between mfcc or plp)
for folder_name in list_sets: # iterating over the list of sets where the features live
print("\nReading dir:", feature_dir + folder_name)
list_ubm_files = util.traverse_dir_2(
ubm_dir,
'*_{}g_{}{}-{}del_{}.dubm'.format(g, feats_info[0],
feats_info[2],
feats_info[1],
ubm_folder_name))
list_mfcc_files = util.traverse_dir_2(
feature_dir + folder_name,
'*{}_{}_{}del.{}'.format(feats_info[0], folder_name,
feats_info[1], feats_info[2]))
extract_fishers.compute_fishers_pretr_ubm_2(
list_mfcc_files=list_mfcc_files,
out_dir=out_dir,
feats_info=feats_info,
list_files_ubm=list_ubm_files,
recipe=recipe,
folder_name=folder_name)
def do_ivecs():
print("=======i-vector extraction phase========")
mfccs_dir = work_dir + 'data/{}/'.format(recipe)
ubm_dir = work_dir + 'data/UBMs/{}'.format(ubm_folder_name)
for deltas in [2]:
feats_info = [20, deltas,
'mfcc'] # info of the mfccs (n_features, deltas)
# the following is a list in the case of the UBM is meant to be trained with training and dev sets
list_files_ubm = [
ubm_dir + '/mfcc_{}_{}_{}del.mfcc'.format(feats_info[0],
ubm_folder_name, deltas)
]
for folder_name in list_sets:
print("\nReading dir:", mfccs_dir + folder_name)
list_mfcc_files = util.traverse_dir_2(
mfccs_dir + folder_name,
'*{}_{}_{}del.{}'.format(feats_info[0], folder_name,
feats_info[1], feats_info[2]))
extract_ivecs.compute_ivecs(list_n_gauss=list_n_clusters,
list_mfcc_files=list_mfcc_files,
out_dir=out_dir,
list_files_ubm=list_files_ubm,
recipe=recipe,
mfcc_info=feats_info,
folder_name=folder_name)
def do_fishers():
print("=======fisher-vector extraction phase========")
feature_dir = work_dir + '/data/{}/'.format(recipe)
for deltas in [0, 1, 2]:
# info-purpose parameters from the frame-level extracted features #
feats_info = [
20, deltas, 'mfcc'
] # info of the features (n_features/dimension, deltas, cepstral_type=choose between mfcc or plp)
obs = ''
# Format is: "featureType_recipeName_nMFCCs_nDeltas.mfcc"
list_files_ubm = [
work_dir +
'data/demencia94ABC/wav16k_split_long/mfcc_demencia94ABC_20_wav16k_split_long_{}del.mfcc'
.format(deltas)
]
for folder_name in list_sets:
list_mfcc_files = util.traverse_dir_2(
feature_dir + folder_name,
'*{}_{}_{}del.{}'.format(feats_info[0], folder_name,
feats_info[1], feats_info[2]))
print(list_mfcc_files)
extract_fishers.compute_fishers(list_n_clusters,
list_mfcc_files,
out_dir,
feats_info=feats_info,
list_files_ubm=list_files_ubm,
recipe=recipe,
folder_name=folder_name)
def do_dimension_reduction():
print("=======dimension reduction phase========")
feature_dir = work_dir + '/data/{}/'.format(recipe)
for delta in [0, 1, 2]:
# info-purpose parameters from the frame-level extracted features #
feats_info = [
40, delta, 'mfcc'
] # info of the features (n_features/dimension, deltas, cepstral_type=choose between mfcc or plp)
obs = '_hires' # observations of the features' config (if there is such) e.g. '_hires' (when the mfccs were extracted using 'hires' params)
list_files_ubm = [
work_dir + '/data/mask/train/{}_mask_{}_train_{}del{}.{}'.format(
feats_info[2], feats_info[0], delta, obs, feats_info[2])
]
pca = pca_trainer(list_files_ubm[0],
n_components=0.97) # train PCA using training set
for folder_name in list_sets:
print("\nReading dir:", feature_dir + folder_name)
list_mfcc_file = util.traverse_dir_2(
feature_dir + folder_name,
'*{}_{}_{}del.{}'.format(feats_info[0], folder_name,
feats_info[1], feats_info[2]))
for item in list_mfcc_file: # transform each dataset
reduced_data = pca_transformer(pca, item)
util.save_pickle(
feature_dir + folder_name + '*{}_{}_{}del{}_pca.{}'.format(
feats_info[0], folder_name, feats_info[1], obs,
feats_info[2]), reduced_data)
def do_fishers():
print("=======fisher-vector extraction phase========")
feature_dir = work_dir + '/data/{}/'.format(recipe)
for delta in [0]:
# info-purpose parameters from the frame-level extracted features #
feats_info = [
23, delta, 'mfcc'
] # info of the features (n_features/dimension, deltas, cepstral_type=choose between mfcc or plp)
obs = '' # observations of the features' config e.g. '_hires' (when the mfccs were extracted using 'hires' params)
list_files_ubm = [
work_dir + '/data/{}/train/{}_{}_{}_train_{}del.{}'.format(
recipe, feats_info[2], recipe, feats_info[0], delta,
feats_info[2])
]
# Format is: "featureType_recipeName_nMFCCs_nDeltas.mfcc"
for folder_name in list_sets:
print("\nReading dir:", feature_dir + folder_name)
list_mfcc_files = util.traverse_dir_2(
feature_dir + folder_name,
'*{}_{}_{}del.{}'.format(feats_info[0], folder_name,
feats_info[1], feats_info[2]))
print(list_mfcc_files)
extract_fishers.compute_fishers(list_n_clusters,
list_mfcc_files,
out_dir,
feats_info=feats_info,
list_files_ubm=list_files_ubm,
recipe=recipe,
folder_name=folder_name)
def do_ivecs():
print("=======i-vector extraction phase========")
mfccs_dir = work_dir + 'data/{}/'.format(recipe)
for deltas in [2]:
feats_info = [20, deltas,
'mfcc'] # info of the mfccs (n_features, deltas)
list_files_ubm = [
work_dir +
'data/demencia94ABC/wav16k_split_long/mfcc_demencia94ABC_20_wav16k_split_long_{}del.mfcc'
.format(deltas)
]
for folder_name in list_sets:
print("\nReading dir:", mfccs_dir + folder_name)
list_mfcc_files = util.traverse_dir_2(
mfccs_dir + folder_name,
'*{}_{}_{}del.{}'.format(feats_info[0], folder_name,
feats_info[1], feats_info[2]))
extract_ivecs.compute_ivecs(list_n_gauss=list_n_clusters,
list_mfcc_files=list_mfcc_files,
out_dir=out_dir,
list_files_ubm=list_files_ubm,
recipe=recipe,
mfcc_info=feats_info,
folder_name=folder_name)
def do_fishers_pretrained_ubm():
mfccs_dir = work_dir + 'data/{}/'.format(recipe)
out_dir = work_dir + 'data/' # Where the computed features will live in
ubm_dir = work_dir + 'data/' + recipe + '/UBMs/' # where the diagonal ubms live
list_ubm_files = util.traverse_dir(
ubm_dir, '.mdl'
) # reading all the files with .mdl or .dubm as format (latter is more reliable)
mfcc_n_deltas = 2 # Number of deltas of the mfccs
list_sets = ['monologue_erlangen', 'readtext_erlangen']
for folder_name in list_sets: # iterating over the list of sets where the features live
print("\nReading dir:", mfccs_dir + folder_name)
for ubm in list_ubm_files: # iterating over the pretrained ubms
list_mfcc_files = util.traverse_dir_2(
mfccs_dir + folder_name, '*{}del.mfcc'.format(
mfcc_n_deltas)) # reading MFCCs to extracting fishers from
extract_fishers.compute_fishers_pretr_ubm(
list_mfcc_files=list_mfcc_files,
out_dir=out_dir,
file_ubm=ubm,
recipe=recipe,
folder_name=folder_name)
feature_dir = work_dir + 'data/{}/'.format(recipe)
out_dir = work_dir + 'data/' # Where the computed features will be saved
ubm_dir = work_dir + 'data/UBMs/' + ubm_folder_name +'/ivec_models/' # where the diagonal ubms live
# list_ubm_files = util.traverse_dir(ubm_dir, '.dubm') # reading all the files with .mdl or .dubm as format (latter is more reliable)
for g in list_n_clusters:
# info-purpose parameters from the frame-level extracted features #
feats_info_ubm = [20, 0, 'mfcc'] # info of the features (n_features/dimension, deltas, cepstral_type=choose between mfcc or plp)
feats_info_flevel = [20, 0, 'fbank'] # info of the features (n_features/dimension, deltas, cepstral_type=choose between mfcc or plp)
for folder_name in list_sets: # iterating over the list of sets where the features live
# print("\nReading dir:", feature_dir + folder_name+'/fisher/')
# here the FULL-diag ubm is used
list_ubm_files = util.traverse_dir_2(ubm_dir, '*_{}g_{}{}-{}del_{}.fubm'.format(g, feats_info_ubm[0],
feats_info_ubm[2],
feats_info_ubm[1],
ubm_folder_name))
list_mfcc_files = util.traverse_dir_2(feature_dir + folder_name+'/', '*{}_{}_{}del.{}'.format(
feats_info_flevel[0],
folder_name,
feats_info_flevel[1],
feats_info_flevel[2]
))
extract_ivecs.extract_ivecs(list_mfcc_files=list_mfcc_files, g=g, list_fubms=list_ubm_files,
mfcc_info=feats_info_ubm, recipe=recipe, out_dir=out_dir, folder_name=folder_name)
fubmfile='/media/jose/hk-data/PycharmProjects/the_speech/data/UBMs/wav16k_split_long/ivec_models/fubm_mdl_256g_20mfcc-0del_wav16k_split_long.fubm'