def test(in_audio_list=None,
lab_dir=None,
results_dir=None,
model_list=None,
model_file=None,
feature_type='PLP_0',
n_coeffs_per_frame=13,
acc_frames=31,
samp_period=0.01,
window_length=0.02):
'''
Test Speech Activity Detection for a list of files given a specific model. Ideally, many
of the input arguments like samp_period, window_length, n_coeffs_per_frame, acc_frames
should be read from the model file.
Input:
in_audio_list : list of audio files (absolute paths)
lab_dir : directory where the .lab transcription files lie
results_dir : directory where the results will be stored
model_list : file containing a list of the class names
model_file : an HTK formatted mmf file (containing the gmm models for the different classes
feature_type : an HTK-formatted string describing the feature_type, e.g., MFCC_0
n_coeffs_per_frame : number of features per frame
acc_frames : number of frames to accumulate features over
samp_period : the frame period (in seconds)
window_length : the frame duration (in seconds)
Output:
conf_matrix : return the confusion matrix
'''
if not os.path.exists(results_dir):
os.makedirs(results_dir)
# Classify using
estimate_channel_gmm(in_audio_list, model_list, model_file, feature_type,
n_coeffs_per_frame, acc_frames, results_dir,
results_dir)
# The annotations are in .lab format, i.e., start_time end_time label per line
ref_annotations_list = os.path.join(results_dir,
'ref_test_annotations.list')
hyp_annotations_list = os.path.join(results_dir,
'hyp_test_annotations.list')
lists.create_corresponding_list_assert(in_audio_list, lab_dir,
ref_annotations_list, 'lab')
lists.create_corresponding_list_assert(in_audio_list, results_dir,
hyp_annotations_list, 'rec')
# Given the results and the reference annotations, evaluate by estimating a confusion matrix
conf_matrix = vad_gmm_evaluate.vad_gmm_evaluate_frames(
ref_annotations_list,
hyp_annotations_list,
samp_period,
model_list,
mode='single')
msg = "{} \n {}".format(model_file, conf_matrix)
logging.info(msg)
return (conf_matrix)
def test(in_audio_list=None, audio_dir=None, ldc_annotations_list=None, lab_dir=None, results_dir=None, model_list=None,
model_file=None, feature_type='PLP_0', n_coeffs_per_frame=13, acc_frames=31,
samp_period=0.01, window_length=0.025, eval_script=None):
'''
Test Speech Activity Detection for a list of files given a specific model. Ideally, many
of the input arguments like samp_period, window_length, n_coeffs_per_frame, acc_frames
should be read from the model file.
Input:
in_audio_list : list of audio files (absolute paths)
audio_dir : directory where the audio files lie
ldc_annotations_list : list with the LDC annotation files
lab_dir : directory where the .lab transcription files lie
results_dir : directory where the results will be stored
model_list : file containing a list of the class names
model_file : an HTK formatted mmf file (containing the gmm models for the different classes
feature_type : an HTK-formatted string describing the feature_type, e.g., MFCC_0
n_coeffs_per_frame : number of features per frame
acc_frames : number of frames to accumulate features over
samp_period : the frame period (in seconds)
window_length : the frame duration (in seconds)
eval_script : the java DARPA evaluation script
Output:
conf_matrix : return the confusion matrix
'''
if not os.path.exists(results_dir):
os.makedirs(results_dir)
# Run the VAD to get the results
vad_gmm.vad_gmm_list(in_audio_list, model_list, model_file, feature_type,
n_coeffs_per_frame, acc_frames, results_dir, results_dir, samp_period, window_length)
# The annotations are in .lab format, i.e., start_time end_time label per line
ref_annotations_list = os.path.join(results_dir,'ref_test_annotations.list')
hyp_annotations_list = os.path.join(results_dir,'hyp_test_annotations.list')
lists.create_corresponding_list_assert(in_audio_list, lab_dir, ref_annotations_list,'lab')
lists.create_corresponding_list_assert(in_audio_list, results_dir, hyp_annotations_list,'rec')
# Given the results and the reference annotations, evaluate by estimating a confusion matrix
conf_matrix = vad_gmm_evaluate.vad_gmm_evaluate_frames(ref_annotations_list, hyp_annotations_list,
samp_period, model_list)
msg = "{0} \n {1}".format(model_file, conf_matrix)
logging.info(msg)
# Estimate accuracy
n_instances = np.sum(conf_matrix)
n_correct = np.sum(conf_matrix.diagonal())
msg = "Accuracy: {0} / {1} = {2}".format(n_correct, n_instances, float(n_correct) / n_instances )
logging.info(msg)
if eval_script is not None and os.path.exists(eval_script):
lists.create_corresponding_list_assert(in_audio_list, lab_dir, ref_annotations_list,'lab.frames.txt')
lists.create_corresponding_list_assert(in_audio_list, results_dir, hyp_annotations_list,'rec.frames.txt')
vad_evaluate_darpa(testing_list=in_audio_list, ref_annotations_list=ldc_annotations_list,
hyp_annotations_list=hyp_annotations_list, eval_script=eval_script,
audio_dir=audio_dir, smp_period=samp_period, window_length=window_length,
results_dir=results_dir, task_id='{0}_{1}_{2}'.format(feature_type, str(n_coeffs_per_frame),
str(acc_frames)))
return conf_matrix
def test(in_audio_list=None, lab_dir=None, results_dir=None, model_list=None,
model_file=None, feature_type='PLP_0', n_coeffs_per_frame=13, acc_frames=31,
samp_period=0.01, window_length=0.02):
'''
Test Speech Activity Detection for a list of files given a specific model. Ideally, many
of the input arguments like samp_period, window_length, n_coeffs_per_frame, acc_frames
should be read from the model file.
Input:
in_audio_list : list of audio files (absolute paths)
lab_dir : directory where the .lab transcription files lie
results_dir : directory where the results will be stored
model_list : file containing a list of the class names
model_file : an HTK formatted mmf file (containing the gmm models for the different classes
feature_type : an HTK-formatted string describing the feature_type, e.g., MFCC_0
n_coeffs_per_frame : number of features per frame
acc_frames : number of frames to accumulate features over
samp_period : the frame period (in seconds)
window_length : the frame duration (in seconds)
Output:
conf_matrix : return the confusion matrix
'''
if not os.path.exists(results_dir):
os.makedirs(results_dir)
# Classify using
estimate_channel_gmm(in_audio_list, model_list, model_file, feature_type,
n_coeffs_per_frame, acc_frames, results_dir, results_dir)
# The annotations are in .lab format, i.e., start_time end_time label per line
ref_annotations_list = os.path.join(results_dir,'ref_test_annotations.list')
hyp_annotations_list = os.path.join(results_dir,'hyp_test_annotations.list')
lists.create_corresponding_list_assert(in_audio_list, lab_dir, ref_annotations_list,'lab')
lists.create_corresponding_list_assert(in_audio_list, results_dir, hyp_annotations_list,'rec')
# Given the results and the reference annotations, evaluate by estimating a confusion matrix
conf_matrix = vad_gmm_evaluate.vad_gmm_evaluate_frames(ref_annotations_list, hyp_annotations_list,
samp_period, model_list, mode='single')
msg = "{} \n {}".format(model_file, conf_matrix)
logging.info(msg)
return(conf_matrix)
def test_gmm_classification(models=None, n_samples=10000, working_dir=os.getcwd(),
apply_hlda=False, hlda_nuisance_dims=2, do_not_generate_features=False):
test_percentage = 0.2
samp_period = 0.01
generate_features = not do_not_generate_features
if not os.path.exists(working_dir):
os.makedirs(working_dir)
fea_dir = os.path.join(working_dir, 'features')
lab_dir = os.path.join(working_dir, 'lab')
list_dir = os.path.join(working_dir, 'lists')
model_dir = os.path.join(working_dir, 'models')
results_dir = os.path.join(working_dir, 'results')
if not os.path.exists(list_dir):
os.makedirs(list_dir)
if not os.path.exists(fea_dir):
os.makedirs(fea_dir)
if not os.path.exists(lab_dir):
os.makedirs(lab_dir)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if not os.path.exists(results_dir):
os.makedirs(results_dir)
# Create a list of the model names
model_list = os.path.join(working_dir,'models.list')
ml = open(model_list,'w')
labels = []
for md in models:
ml.write('{}\n'.format(md.name))
labels.append(md.name)
ml.close()
tst_file_list = os.path.join(list_dir,'test_files.list')
trn_file_list = os.path.join(list_dir,'train_files.list')
if generate_features:
fea_file_names = write_feature_files(models, n_samples,
fea_dir, lab_dir)
n_fea_files = len(fea_file_names)
logging.info('Number of feature files: {}'.format(n_fea_files))
n_train_files = int(round(n_fea_files*(1-test_percentage)))
trn_list = open(trn_file_list,'w')
for f_n in fea_file_names[0:n_train_files]:
trn_list.write("{}\n".format(os.path.join(fea_dir,f_n)))
trn_list.close()
tst_list = open(tst_file_list,'w')
for f_n in fea_file_names[n_train_files:n_fea_files-1]:
tst_list.write("{}\n".format(os.path.join(fea_dir,f_n)))
tst_list.close()
n_mixes = models[0].n_components
n_dims = models[0].means.shape[1]
train_gmm_classifier.train_M_sized_gmm_classifier(n_mixes, n_dims, 'USER', trn_file_list,
model_list, lab_dir, model_dir, apply_hlda,
hlda_nuisance_dims)
model_file = os.path.join(model_dir,'newMacros')
test_gmm_classifier.test_gmm_classifier(tst_file_list, model_list, model_file, results_dir)
ref_annotations_list = os.path.join(list_dir,'labels.list')
hyp_annotations_list = os.path.join(list_dir,'hyp.list')
lists.create_corresponding_list_assert(tst_file_list,lab_dir,ref_annotations_list,'lab')
lists.create_corresponding_list_assert(tst_file_list,results_dir,hyp_annotations_list,'rec')
conf_matrix = evaluate_gmm_classifier.evaluate_results_list(ref_annotations_list,
hyp_annotations_list,
samp_period, labels)
logging.info('Confusion matrix: {}'.format(conf_matrix))
def test(in_audio_list=None, audio_dir=None, ldc_annotations_list=None, lab_dir=None, results_dir=None, model_list=None,
model_file=None, feature_type='PLP_0', n_coeffs_per_frame=13, acc_frames=31,
samp_period=0.01, window_length=0.025, eval_script=None):
'''
Test Speech Activity Detection for a list of files given a specific model. Ideally, many
of the input arguments like samp_period, window_length, n_coeffs_per_frame, acc_frames
should be read from the model file.
Input:
in_audio_list : list of audio files (absolute paths)
audio_dir : directory where the audio files lie
ldc_annotations_list : list with the LDC annotation files
lab_dir : directory where the .lab transcription files lie
results_dir : directory where the results will be stored
model_list : file containing a list of the class names
model_file : an HTK formatted mmf file (containing the gmm models for the different classes
feature_type : an HTK-formatted string describing the feature_type, e.g., MFCC_0
n_coeffs_per_frame : number of features per frame
acc_frames : number of frames to accumulate features over
samp_period : the frame period (in seconds)
window_length : the frame duration (in seconds)
eval_script : the java DARPA evaluation script
Output:
conf_matrix : return the confusion matrix
'''
if not os.path.exists(results_dir):
os.makedirs(results_dir)
# Run the VAD to get the results
vad_gmm.vad_gmm_list(in_audio_list, model_list, model_file, feature_type,
n_coeffs_per_frame, acc_frames, results_dir, results_dir, samp_period, window_length)
# The annotations are in .lab format, i.e., start_time end_time label per line
ref_annotations_list = os.path.join(results_dir, 'ref_test_annotations.list')
hyp_annotations_list = os.path.join(results_dir, 'hyp_test_annotations.list')
lists.create_corresponding_list_assert(in_audio_list, lab_dir, ref_annotations_list,'lab')
lists.create_corresponding_list_assert(in_audio_list, results_dir, hyp_annotations_list,'rec')
# Given the results and the reference annotations, evaluate by estimating a confusion matrix
conf_matrix = vad_gmm_evaluate.vad_gmm_evaluate_frames(ref_annotations_list, hyp_annotations_list,
samp_period, model_list)
msg = "{} \n {}".format(model_file, conf_matrix)
logging.info(msg)
# Estimate accuracy
n_instances = np.sum(conf_matrix)
n_correct = np.sum(conf_matrix.diagonal())
msg = "Accuracy: {} / {} = {}".format(n_correct, n_instances, float(n_correct) / n_instances )
logging.info(msg)
if eval_script is not None and os.path.exists(eval_script):
lists.create_corresponding_list_assert(in_audio_list, lab_dir, ref_annotations_list,'lab.frames.txt')
lists.create_corresponding_list_assert(in_audio_list, results_dir, hyp_annotations_list,'rec.frames.txt')
vad_evaluate_darpa(testing_list=in_audio_list, ref_annotations_list=ldc_annotations_list,
hyp_annotations_list=hyp_annotations_list, eval_script=eval_script,
audio_dir=audio_dir, smp_period=samp_period, window_length=window_length,
results_dir=results_dir, task_id='{}_{}_{}'.format(feature_type, str(n_coeffs_per_frame),
str(acc_frames)))
return(conf_matrix)
def test_gmm_classification(models=None,
n_samples=10000,
working_dir=os.getcwd(),
apply_hlda=False,
hlda_nuisance_dims=2,
do_not_generate_features=False):
test_percentage = 0.2
samp_period = 0.01
generate_features = not do_not_generate_features
if not os.path.exists(working_dir):
os.makedirs(working_dir)
fea_dir = os.path.join(working_dir, 'features')
lab_dir = os.path.join(working_dir, 'lab')
list_dir = os.path.join(working_dir, 'lists')
model_dir = os.path.join(working_dir, 'models')
results_dir = os.path.join(working_dir, 'results')
if not os.path.exists(list_dir):
os.makedirs(list_dir)
if not os.path.exists(fea_dir):
os.makedirs(fea_dir)
if not os.path.exists(lab_dir):
os.makedirs(lab_dir)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if not os.path.exists(results_dir):
os.makedirs(results_dir)
# Create a list of the model names
model_list = os.path.join(working_dir, 'models.list')
ml = open(model_list, 'w')
labels = []
for md in models:
ml.write('{}\n'.format(md.name))
labels.append(md.name)
ml.close()
tst_file_list = os.path.join(list_dir, 'test_files.list')
trn_file_list = os.path.join(list_dir, 'train_files.list')
if generate_features:
fea_file_names = write_feature_files(models, n_samples, fea_dir,
lab_dir)
n_fea_files = len(fea_file_names)
logging.info('Number of feature files: {}'.format(n_fea_files))
n_train_files = int(round(n_fea_files * (1 - test_percentage)))
trn_list = open(trn_file_list, 'w')
for f_n in fea_file_names[0:n_train_files]:
trn_list.write("{}\n".format(os.path.join(fea_dir, f_n)))
trn_list.close()
tst_list = open(tst_file_list, 'w')
for f_n in fea_file_names[n_train_files:n_fea_files - 1]:
tst_list.write("{}\n".format(os.path.join(fea_dir, f_n)))
tst_list.close()
n_mixes = models[0].n_components
n_dims = models[0].means.shape[1]
train_gmm_classifier.train_M_sized_gmm_classifier(n_mixes, n_dims, 'USER',
trn_file_list,
model_list, lab_dir,
model_dir, apply_hlda,
hlda_nuisance_dims)
model_file = os.path.join(model_dir, 'newMacros')
test_gmm_classifier.test_gmm_classifier(tst_file_list, model_list,
model_file, results_dir)
ref_annotations_list = os.path.join(list_dir, 'labels.list')
hyp_annotations_list = os.path.join(list_dir, 'hyp.list')
lists.create_corresponding_list_assert(tst_file_list, lab_dir,
ref_annotations_list, 'lab')
lists.create_corresponding_list_assert(tst_file_list, results_dir,
hyp_annotations_list, 'rec')
conf_matrix = evaluate_gmm_classifier.evaluate_results_list(
ref_annotations_list, hyp_annotations_list, samp_period, labels)
logging.info('Confusion matrix: {}'.format(conf_matrix))
