def vad_hmm_train(audio_list,
annotations_dir,
model_list,
feature_type,
n_coeffs_per_frame,
acc_frames,
working_dir,
apply_hlda=False,
hlda_nuisance_dims=0,
n_states=10,
n_gmm_comps=None,
acc_frame_shift=1,
n_train_iterations=10,
samp_period=0.01):
logging.info('Number of hmm states: ' + str(n_gmm_comps))
features_dir = os.path.join(working_dir, feature_type)
acc_features_dir = features_dir + '_frames'
fea_file_list = os.path.join(working_dir, 'feature_files.list')
if not os.path.exists(features_dir):
os.makedirs(features_dir)
# Feature extraction
vef.fea_extract(audio_list, feature_type, n_coeffs_per_frame, features_dir,
samp_period)
vef.create_corresponding_list_assert(audio_list, features_dir,
fea_file_list, 'fea')
if not os.path.exists(acc_features_dir):
os.makedirs(acc_features_dir)
vef.accumulate_feature_vectors_parallel(fea_file_list, acc_frames,
acc_features_dir, acc_frame_shift)
vef.create_corresponding_list_assert(audio_list, acc_features_dir,
fea_file_list, 'fea')
# GMM training
if n_gmm_comps == None:
raise ValueError(
'Automatic determination of the number of the GMM components not implemented'
)
model_dir = os.path.join(working_dir, 'models')
train_gmm_classifier.train_hmm_classifier_incrementally(
n_gmm_comps, n_states, n_coeffs_per_frame, feature_type, fea_file_list,
model_list, annotations_dir, model_dir, apply_hlda, hlda_nuisance_dims,
n_train_iterations)
model_files = []
m_f = os.path.join(model_dir, 'newMacros')
model_files.append(m_f)
m_f_no_hlda = m_f + "_no_hlda"
model_files.append(m_f_no_hlda)
assert (os.path.exists(m_f))
return model_files
def vad_gmm_list(audio_list, model_list, model_file, feature_type,
n_coeffs_per_frame, acc_frames, results_dir, working_dir, samp_period, win_length):
features_dir = os.path.join(working_dir,feature_type)
acc_features_dir = features_dir+'_frames'
fea_file_list = os.path.join(working_dir,'feature_files.list')
if not os.path.exists(features_dir):
os.makedirs(features_dir)
if not os.path.exists(acc_features_dir):
os.makedirs(acc_features_dir)
vef.fea_extract(audio_list,feature_type,n_coeffs_per_frame,features_dir,samp_period, win_length)
vef.create_corresponding_list_assert(audio_list,features_dir,fea_file_list,'fea')
vef.accumulate_feature_vectors(fea_file_list,acc_frames,acc_features_dir)
vef.create_corresponding_list_assert(audio_list,acc_features_dir,fea_file_list,'fea')
test_gmm_classifier(fea_file_list, model_list, model_file, results_dir)
def estimate_channel_gmm(audio_list, model_list, model_file, feature_type,
n_coeffs_per_frame, acc_frames, results_dir, working_dir):
features_dir = os.path.join(working_dir,feature_type)
acc_features_dir = features_dir+'_frames'
fea_file_list = os.path.join(working_dir,'feature_files.list')
if not os.path.exists(features_dir):
os.makedirs(features_dir)
if not os.path.exists(acc_features_dir):
os.makedirs(acc_features_dir)
vef.fea_extract(audio_list,feature_type,n_coeffs_per_frame,features_dir)
vef.create_corresponding_list_assert(audio_list,features_dir,fea_file_list,'fea')
vef.accumulate_feature_vectors(fea_file_list,acc_frames,acc_features_dir)
vef.create_corresponding_list_assert(audio_list,acc_features_dir,fea_file_list,'fea')
test_gmm_classifier.test_gmm_classifier(test_list=fea_file_list, model_list=model_list,
model_file=model_file, results_dir=results_dir,
mode='single')
def vad_gmm_list(audio_list, model_list, model_file, feature_type,
n_coeffs_per_frame, acc_frames, results_dir, working_dir,
samp_period, win_length):
features_dir = os.path.join(working_dir, feature_type)
acc_features_dir = features_dir + '_frames'
fea_file_list = os.path.join(working_dir, 'feature_files.list')
if not os.path.exists(features_dir):
os.makedirs(features_dir)
if not os.path.exists(acc_features_dir):
os.makedirs(acc_features_dir)
vef.fea_extract(audio_list, feature_type, n_coeffs_per_frame, features_dir,
samp_period, win_length)
vef.create_corresponding_list_assert(audio_list, features_dir,
fea_file_list, 'fea')
vef.accumulate_feature_vectors(fea_file_list, acc_frames, acc_features_dir)
vef.create_corresponding_list_assert(audio_list, acc_features_dir,
fea_file_list, 'fea')
test_gmm_classifier(fea_file_list, model_list, model_file, results_dir)
def vad_gmm_train(audio_list, annotations_dir, model_list, feature_type,
n_coeffs_per_frame, acc_frames, working_dir, apply_hlda=False,
hlda_nuisance_dims=0, n_gmm_comps=None, acc_frame_shift=1,
n_train_iterations=10, samp_period=0.01, win_length=0.025, segment_boundaries=[1/3.0, 2/3.0]):
logging.info('Number of gmm components: '+str(n_gmm_comps))
features_dir = os.path.join(working_dir,feature_type)
acc_features_dir = features_dir+'_frames'
fea_file_list = os.path.join(working_dir,'feature_files.list')
if not os.path.exists(features_dir):
os.makedirs(features_dir)
# Feature extraction
vef.fea_extract(audio_list, feature_type, n_coeffs_per_frame, features_dir, samp_period, win_length)
vef.create_corresponding_list_assert(audio_list, features_dir, fea_file_list, 'fea')
if not os.path.exists(acc_features_dir):
os.makedirs(acc_features_dir)
vef.accumulate_feature_vectors_parallel(fea_file_list, acc_frames,
acc_features_dir, acc_frame_shift)
vef.create_corresponding_list_assert(audio_list,acc_features_dir,fea_file_list,'fea')
#vef.split_feature_files_labels(fea_file_list, annotations_dir, label_list,
# boundaries=segment_boundaries)
# GMM training
if n_gmm_comps==None:
raise ValueError('Automatic determination of the number of the GMM components not implemented')
model_dir = os.path.join(working_dir,'models')
train_gmm_classifier.train_M_sized_gmm_classifier_incrementally(n_gmm_comps, n_coeffs_per_frame,
feature_type, fea_file_list, model_list,
annotations_dir, model_dir, apply_hlda,
hlda_nuisance_dims, n_train_iterations)
model_files = []
m_f = os.path.join(model_dir,'newMacros')
model_files.append(m_f)
m_f_no_hlda = m_f+"_no_hlda"
model_files.append(m_f_no_hlda)
assert(os.path.exists(m_f))
return model_files
def estimate_channel_gmm(audio_list, model_list, model_file, feature_type,
n_coeffs_per_frame, acc_frames, results_dir,
working_dir):
features_dir = os.path.join(working_dir, feature_type)
acc_features_dir = features_dir + '_frames'
fea_file_list = os.path.join(working_dir, 'feature_files.list')
if not os.path.exists(features_dir):
os.makedirs(features_dir)
if not os.path.exists(acc_features_dir):
os.makedirs(acc_features_dir)
vef.fea_extract(audio_list, feature_type, n_coeffs_per_frame, features_dir)
vef.create_corresponding_list_assert(audio_list, features_dir,
fea_file_list, 'fea')
vef.accumulate_feature_vectors(fea_file_list, acc_frames, acc_features_dir)
vef.create_corresponding_list_assert(audio_list, acc_features_dir,
fea_file_list, 'fea')
test_gmm_classifier.test_gmm_classifier(test_list=fea_file_list,
model_list=model_list,
model_file=model_file,
results_dir=results_dir,
mode='single')
