def load_norm_ivecs(self):
""" Load normalization i-vectors, scale and shift files and also pretrained model.
:returns: i-vectors
:rtype: numpy.array
"""
line = None
with open(self.norm_list, 'r') as f:
for line in f:
line = line.rstrip()
loginfo(
'[Diarization.load_norm_ivecs] Loading npy file {} ...'.
format(line))
try:
yield np.load('{}.npy'.format(
os.path.join(self.ivecs_dir, line))).flatten()
except IOError:
logwarning(
'[Diarization.load_norm_ivecs] No pickle file found for {}.'
.format(line))
self.scale = np.load(
os.path.join(self.ivecs_dir, os.path.dirname(line), 'scale.npy'))
self.shift = np.load(
os.path.join(self.ivecs_dir, os.path.dirname(line), 'shift.npy'))
try:
with open(
os.path.join(self.ivecs_dir, os.path.dirname(line),
'model.pkl')) as f:
self.model = pickle.load(f)
except IOError:
logwarning(
'[Diarization.load_norm_ivecs] No pretrained model found.')
def dump_rttm(self, scores):
for ivecset in self.ivecs:
if ivecset.size() > 0:
name = ivecset.name
# dirty trick, will be removed, watch out
if 'beamformed' in ivecset.name:
ivecset.name = re.sub('beamformed/', '', ivecset.name)
# # # # # # # # # # # # # # # # # # # # #
reg_name = re.sub('/.*', '', ivecset.name)
Tools.mkdir_p(os.path.join(self.out_dir,
os.path.dirname(name)))
with open(os.path.join(self.out_dir, name + '.rttm'),
'w') as f:
for i, ivec in enumerate(ivecset.ivecs):
start, end = ivec.window_start, ivec.window_end
idx = np.argmax(scores[name].T[i])
f.write(
'SPEAKER {} 1 {} {} <NA> <NA> {}_spkr_{} <NA>\n'.
format(reg_name, float(start / 1000.0),
float((end - start) / 1000.0), reg_name,
idx))
else:
logwarning(
'[Diarization.dump_rttm] No i-vectors to dump in {}.'.
format(ivecset.name))
def score(self):
scores_dict = {}
for ivecset in self.ivecs:
name = os.path.normpath(ivecset.name)
ivecs = ivecset.get_all()
loginfo('[Diarization.score] Scoring {} ...'.format(name))
size = ivecset.size()
if size > 0:
if ivecset.num_speakers is not None:
num_speakers = min(ivecset.num_speakers, size)
sklearnkmeans = sklearnKMeans(
n_clusters=num_speakers).fit(ivecs)
centroids = KMeans(sklearnkmeans.cluster_centers_,
num_speakers, self.plda).fit(ivecs)
else:
num_speakers, centroids = self.get_num_speakers(ivecs)
if self.norm_list is None:
scores_dict[name] = self.plda.score(
ivecs, centroids, self.scale, self.shift)
else:
scores_dict[name] = self.s_norm(ivecs, centroids)
else:
logwarning(
'[Diarization.score] No i-vectors to score in {}.'.format(
ivecset.name))
return scores_dict
def load_ivecs(self):
with open(self.input_list, 'r') as f:
for line in f:
loginfo('[Diarization.load_ivecs] Loading pickle file {} ...'.
format(line.rstrip().split()[0]))
line = line.rstrip()
try:
if len(line.split()) == 1:
with open(os.path.join(self.ivecs_dir,
line + '.pkl')) as i:
yield pickle.load(i)
elif len(line.split()) == 2:
file_name = line.split()[0]
num_spks = int(line.split()[1])
with open(
os.path.join(self.ivecs_dir,
file_name + '.pkl')) as i:
ivec_set = pickle.load(i)
ivec_set.num_speakers = num_spks
yield ivec_set
else:
raise DiarizationException(
'[Diarization.load_ivecs] Unexpected number of columns in input list {}.'
.format(self.input_list))
except IOError:
logwarning(
'[Diarization.load_ivecs] No pickle file found for {}.'
.format(line.rstrip().split()[0]))
def get_der(self, ref_file, scores):
""" Compute Diarization Error Rate from reference and scores.
:param ref_file: path to file with diarization reference
:type ref_file: str
:param scores: input scores from PLDA model
:type scores: numpy.array
"""
ref, hyp = self.init_annotations()
with open(ref_file, 'r') as f:
for line in f:
_, name, _, start, duration, _, _, speaker, _ = line.split()
ref[name][Segment(float(start),
float(start) + float(duration))] = speaker
for ivecset in self.ivecs:
if ivecset.size() > 0:
name, reg_name = ivecset.name, ivecset.name
# dirty trick, will be removed, watch out
if 'beamformed' in name:
reg_name = re.sub('beamformed/', '', name)
# # # # # # # # # # # # # # # # # # # # #
reg_name = re.sub('/.*', '', reg_name)
for i, ivec in enumerate(ivecset.ivecs):
start, end = ivec.window_start / 1000.0, ivec.window_end / 1000.0
hyp[reg_name][Segment(start,
end)] = np.argmax(scores[name].T[i])
else:
logwarning(
'[Diarization.get_der] No i-vectors to dump in {}.'.format(
ivecset.name))
der = DiarizationErrorRate()
der.collar = 0.25
names, values, summ = [], [], 0.0
for name in ref.keys():
names.append(name)
der_num = der(ref[name], hyp[name]) * 100
values.append(der_num)
summ += der_num
loginfo('[Diarization.get_der] {} DER = {}'.format(
name, '{0:.3f}'.format(der_num)))
loginfo('[Diarization.get_der] Average DER = {}'.format(
'{0:.3f}'.format(summ / float(len(ref.keys())))))
Diarization.plot_der(names, values)
def score(self):
""" Score i-vectors agains speaker clusters.
:returns: PLDA scores
:rtype: numpy.array
"""
scores_dict = {}
for ivecset in self.ivecs:
name = os.path.normpath(ivecset.name)
ivecs = ivecset.get_all()
loginfo('[Diarization.score] Scoring {} ...'.format(name))
size = ivecset.size()
if size > 0:
if ivecset.num_speakers is not None:
num_speakers = ivecset.num_speakers
sklearnkmeans = sklearnKMeans(
n_clusters=num_speakers).fit(ivecs)
centroids = KMeans(sklearnkmeans.cluster_centers_,
num_speakers, self.plda).fit(ivecs)
else:
if self.norm_ivecs is not None:
num_speakers, centroids = self.get_num_speakers(ivecs)
else:
raise DiarizationException(
'[Diarization.score] Can not estimate number of speakers without training set.'
)
if self.norm_list is None:
scores_dict[name] = self.plda.score(
ivecs, centroids, self.scale, self.shift)
else:
scores_dict[name] = self.s_norm(ivecs, centroids)
else:
logwarning(
'[Diarization.score] No i-vectors to score in {}.'.format(
ivecset.name))
return scores_dict