def read_args(args):
segments, labels = read_labels_file(args.input_label_file)
xvec_all = dict(kaldi_io.read_vec_flt_ark(args.xvector_ark_file))
xvectors = []
for segment in segments:
xvectors.append(xvec_all[segment])
_, _, plda_psi = kaldi_io.read_plda(args.plda)
return xvectors, segments, labels, plda_psi
def read_args(args):
segments, labels = read_labels_file(args.input_label_file)
xvec_all = dict(kaldi_io.read_vec_flt_ark(args.xvector_ark_file))
xvectors = []
for segment in segments:
xvectors.append(xvec_all[segment])
_, _, plda_psi = kaldi_io.read_plda(args.plda)
if (args.overlap_rttm is not None):
print('Getting overlap segments...')
overlaps = get_overlap_vector(args.overlap_rttm, segments)
else:
overlaps = None
return xvectors, segments, labels, plda_psi, overlaps
sys.argv[12]
) # For VB-HMM, x-vectors are reduced to this dimensionality using LDA
Fa = float(
sys.argv[13]) # Parameter of VB-HMM (see VB_diarization.VB_diarization)
Fb = float(
sys.argv[14]) # Parameter of VB-HMM (see VB_diarization.VB_diarization)
LoopP = float(
sys.argv[15]) # Parameter of VB-HMM (see VB_diarization.VB_diarization)
use_VB = True # False for using only AHC
frm_shift = 0.01 # frame rate of MFCC features
glob_tran = kaldi_io.read_mat(
tran_mat_file) # x-vector whitening transformation
glob_mean = kaldi_io.read_vec_flt(mean_vec_file) # x-vector centering vector
kaldi_plda_train = kaldi_io.read_plda(plda_file) # out-of-domain PLDA model
kaldi_plda_adapt = kaldi_io.read_plda(
plda_adapt_file) # in-domain "adaptation" PLDA model
segs_dict = read_xvector_timing_dict(
segments_file) # segments file with x-vector timing information
plda_train_mu, plda_train_tr, plda_train_psi = kaldi_plda_train
plda_adapt_mu, plda_adapt_tr, plda_adapt_psi = kaldi_plda_adapt
# Interpolate across-class, within-class and means of the two PLDA models with interpolation factor "alpha"
plda_mu = alpha * plda_train_mu + (1.0 - alpha) * plda_adapt_mu
W_train = np.linalg.inv(plda_train_tr.T.dot(plda_train_tr))
B_train = np.linalg.inv((plda_train_tr.T / plda_train_psi).dot(plda_train_tr))
W_adapt = np.linalg.inv(plda_adapt_tr.T.dot(plda_adapt_tr))
B_adapt = np.linalg.inv((plda_adapt_tr.T / plda_adapt_psi).dot(plda_adapt_tr))
W = alpha * W_train + (1.0 - alpha) * W_adapt