test_ind_nonsmokers = np.sort(random.sample(range(num_nonsmokers), int(num_smokers / folds)))
train_ind_nonsmokers = np.setdiff1d(np.arange(num_nonsmokers), test_ind_nonsmokers)[0:len(train_ind_smokers)]
# collect training features and labels for smokers and training the smoker model
training_features_smokers = np.array([])
label_smokers = []
label_smokers_numeric = []
for ixs in train_ind_smokers:
spx = 0 # to choose only one recording for each speaker
for jxs in [x for x in os.listdir(input_folder_plp + 'Y/' + smokers[ixs]) if x.endswith('.npy')][:-1]:
# if spx < 5:
plp_features = np.load(input_folder_plp + 'Y/' + smokers[ixs] + '/' + jxs)
if not np.isnan(np.sum(plp_features)):
plp_mean = np.mean(plp_features[:,0:13],axis=0)
plp_std = np.std(plp_features[:,0:13],axis=0,ddof=1)
acoustic_features = a2v(np.concatenate((plp_mean,plp_std),axis=0)).T
# acoustic_features = a2v(plp_features).T
if len(training_features_smokers) == 0:
training_features_smokers = acoustic_features
else:
training_features_smokers = np.append(training_features_smokers,acoustic_features,axis=0)
spx += 1
label_smokers.append('Y')
label_smokers_numeric.append(1)
label_smokers_numeric = np.array(label_smokers_numeric)
# collect training features and labels for non-smokers and training the non-smoker model
training_features_nonsmokers = np.array([])
label_nonsmokers = []
label_nonsmokers_numeric = []
usecols=['speaker ID']).T.values.tolist()[0])
trn_labels = np.asarray(
pd.read_csv(path_training_filelists + training_filelists[itx],
usecols=['smoking label']).T.values.tolist()[0])
#######################################################################
training_features_mfcc = np.array([])
trn_label_numeric = []
for rx in range(len(trn_filenames)):
feat_file_name = path_mfcc_features + trn_labels[rx] + '/' + str(
trn_spk_ids[rx]) + '/' + str(
trn_spk_ids[rx]) + '_' + trn_filenames[rx][:-4] + '.npy'
if os.path.isfile(feat_file_name): # check if the feature file exists
mfcc_features = np.load(feat_file_name) # load MFCCs
mfcc_mean = np.mean(mfcc_features, axis=0) # average over frames
acoustic_features_mfcc = a2v(
mfcc_mean
).T # convert it to vector to be able to append to other features (of other training recordings)
if len(training_features_mfcc
) == 0: # when the array is empty (first time appending)
training_features_mfcc = acoustic_features_mfcc
else: # when the array is not empty
training_features_mfcc = np.append(training_features_mfcc,
acoustic_features_mfcc,
axis=0)
if trn_labels[rx] == 'Y': # make numeric labels (Y:smoker --> 1)
trn_label_numeric.append(1)
elif trn_labels[
rx] == 'N': # make numeric labels (N:nonsmoker --> 2)
trn_label_numeric.append(2)
else:
print('no feature file name -- training -- %d, %d' % (itx, rx))
for ixs in train_ind_smokers:
spx = 0 # to choose only one recording for each speaker
for jxs in [
x for x in os.listdir(input_folder_mfcc + 'Y/' +
smokers[ixs]) if x.endswith('.npy')
][:-1]:
# if spx < 5:
pros_feat_file_name = input_folder_form + 'Y/' + jxs
if os.path.isfile(pros_feat_file_name):
formant_features = np.load(input_folder_form + 'Y/' + jxs)
if not np.isnan(np.sum(formant_features)):
mfcc_features = np.load(input_folder_mfcc + 'Y/' +
smokers[ixs] + '/' + jxs)
mfcc_features = np.mean(mfcc_features, axis=0)
# acoustic_features = a2v(np.concatenate((mfcc_features,formant_features[1:]),axis=0)).T
acoustic_features = a2v(mfcc_features).T
if len(training_features_smokers) == 0:
training_features_smokers = acoustic_features
else:
training_features_smokers = np.append(
training_features_smokers,
acoustic_features,
axis=0)
spx += 1
label_smokers.append('Y')
label_smokers_numeric.append(1)
label_smokers_numeric = np.array(label_smokers_numeric)
# collect training features and labels for non-smokers and training the non-smoker model
training_features_nonsmokers = np.array([])
for ixs in train_ind_smokers:
spx = 0 # to choose only one recording for each speaker
for jxs in [
x
for x in os.listdir(input_folder_plp + 'Y/' + smokers[ixs])
if x.endswith('.npy')
][:-1]:
# if spx < 5:
plp_features = np.load(input_folder_plp + 'Y/' + smokers[ixs] +
'/' + jxs)
plp_mean = np.mean(plp_features[:, 0:13], axis=0)
mfcc_features = np.load(input_folder_mfcc + 'Y/' +
smokers[ixs] + '/' + jxs)
mfcc_mean = np.mean(mfcc_features, axis=0)
# acoustic_features_plp = a2v(np.concatenate((plp_mean,plp_std),axis=0)).T
acoustic_features_plp = a2v(plp_mean).T
acoustic_features_mfcc = a2v(mfcc_mean).T
if len(training_features_smokers_plp) == 0:
training_features_smokers_plp = acoustic_features_plp
training_features_smokers_mfcc = acoustic_features_mfcc
else:
training_features_smokers_plp = np.append(
training_features_smokers_plp,
acoustic_features_plp,
axis=0)
training_features_smokers_mfcc = np.append(
training_features_smokers_mfcc,
acoustic_features_mfcc,
axis=0)
spx += 1
label_smokers.append('Y')
if x.endswith('.npy')
][:-1]:
# if spx < 5:
pros_feat_file_name = input_folder_form + 'Y/' + jxs
if os.path.isfile(pros_feat_file_name):
formant_features = np.load(input_folder_form + 'Y/' + jxs)
if not np.isnan(np.sum(formant_features)):
# pitch_harmonicity = np.load(pros_feat_file_name).T
# pitch_harmonicity[pitch_harmonicity == 0] = np.nan
# pitch_harmonicity = np.nanmean(pitch_harmonicity, axis=0)
# pitch_harmonicity = (pitch_harmonicity - np.tile(np.mean(pitch_harmonicity, axis=0),(pitch_harmonicity.shape[0], 1))) / np.tile(np.std(pitch_harmonicity, axis=0, ddof=1), (pitch_harmonicity.shape[0], 1))
mfcc_features = np.load(input_folder_mfcc + 'Y/' +
smokers[ixs] + '/' + jxs)
mfcc_features = np.mean(mfcc_features, axis=0)
acoustic_features = a2v(
np.concatenate((mfcc_features, formant_features),
axis=0)).T
if len(training_features_smokers) == 0:
training_features_smokers = acoustic_features
else:
training_features_smokers = np.append(
training_features_smokers,
acoustic_features,
axis=0)
spx += 1
label_smokers.append('Y')
label_smokers_numeric.append(1)
label_smokers_numeric = np.array(label_smokers_numeric)
toc_feat_s = time.time()
print(' Done! ---> Elapsed Time: %2.2f sec.' % (toc_feat_s - tic_feat_s))