def __init__(self, spk_id, model, pre_model_dir="pre-models", threshold=0.0):
self.pre_model_dir = os.path.abspath(pre_model_dir)
self.spk_id = os.path.abspath(spk_id)
if not os.path.exists(self.spk_id):
os.makedirs(self.spk_id)
self.audio_dir = os.path.abspath(self.spk_id + "/audio")
self.mfcc_dir = os.path.abspath(self.spk_id + "/mfcc")
self.log_dir = os.path.abspath(self.spk_id + "/log")
self.ivector_dir = os.path.abspath(self.spk_id + "/ivector")
self.threshold = threshold
self.utt_id = model[1]
self.identity_location = model[2]
self.z_norm_mean = model[3]
self.z_norm_std = model[4]
self.train_ivector_scp = self.spk_id + "/ivector.scp"
np.savetxt(self.train_ivector_scp, np.concatenate((np.array([self.utt_id])[:, np.newaxis], np.array([self.identity_location])[:, np.newaxis]), axis=1), fmt="%s")
self.kaldi_helper = ivector_PLDA_kaldiHelper(pre_model_dir=self.pre_model_dir, audio_dir=self.audio_dir,
mfcc_dir=self.mfcc_dir, log_dir=self.log_dir,
ivector_dir=self.ivector_dir)
def __init__(self,
group_id,
model_list,
pre_model_dir="pre-models",
threshold=0.0):
self.pre_model_dir = os.path.abspath(pre_model_dir)
self.group_id = os.path.abspath(group_id)
if not os.path.exists(self.group_id):
os.makedirs(self.group_id)
self.audio_dir = os.path.abspath(self.group_id + "/audio")
self.mfcc_dir = os.path.abspath(self.group_id + "/mfcc")
self.log_dir = os.path.abspath(self.group_id + "/log")
self.ivector_dir = os.path.abspath(self.group_id + "/ivector")
self.threshold = threshold
self.n_speakers = len(model_list)
self.spk_ids = []
self.utt_ids = []
self.identity_locations = []
self.z_norm_means = np.zeros(self.n_speakers, dtype=np.float64)
self.z_norm_stds = np.zeros(self.n_speakers, dtype=np.float64)
for i, model in enumerate(model_list):
spk_id = model[0]
utt_id = model[1]
identity_location = model[2]
mean = model[3]
std = model[4]
self.spk_ids.append(spk_id)
self.utt_ids.append(utt_id)
self.identity_locations.append(identity_location)
self.z_norm_means[i] = mean
self.z_norm_stds[i] = std
''' make sure self.ids is in order, otherwise kaldi may oder them, which may leads to wrong results
'''
self.spk_ids, self.utt_ids, self.identity_locations, self.z_norm_means, self.z_norm_stds = \
self.order(self.spk_ids, self.utt_ids, self.identity_locations, self.z_norm_means, self.z_norm_stds)
self.train_ivector_scp = self.group_id + "/ivector.scp"
np.savetxt(self.train_ivector_scp,
np.concatenate(
(np.array(self.utt_ids)[:, np.newaxis],
np.array(self.identity_locations)[:, np.newaxis]),
axis=1),
fmt="%s")
self.kaldi_helper = ivector_PLDA_kaldiHelper(
pre_model_dir=self.pre_model_dir,
audio_dir=self.audio_dir,
mfcc_dir=self.mfcc_dir,
log_dir=self.log_dir,
ivector_dir=self.ivector_dir)
def score(self, audio_list, fs=16000, bits_per_sample=16, n_jobs=10, debug=False):
if os.path.exists(self.audio_dir):
shutil.rmtree(self.audio_dir)
if os.path.exists(self.mfcc_dir):
shutil.rmtree(self.mfcc_dir)
if os.path.exists(self.log_dir):
shutil.rmtree(self.log_dir)
if os.path.exists(self.ivector_dir):
shutil.rmtree(self.ivector_dir)
if not os.path.exists(self.audio_dir):
os.makedirs(self.audio_dir)
if not os.path.exists(self.mfcc_dir):
os.makedirs(self.mfcc_dir)
if not os.path.exists(self.log_dir):
os.makedirs(self.log_dir)
if not os.path.exists(self.ivector_dir):
os.makedirs(self.ivector_dir)
if isinstance(audio_list, np.ndarray):
if len(audio_list.shape) == 1 or (len(audio_list.shape) == 2 and (audio_list.shape[0] == 1 or audio_list.shape[1] == 1)):
audio_list = [audio_list]
else:
audio_list = [audio_list[:, i] for i in range(audio_list.shape[1])]
else:
audio_list = copy.deepcopy(audio_list) # avoid influencing
for i, audio in enumerate(audio_list):
if not audio.dtype == np.int16:
audio_list[i] = (audio * (2 ** (bits_per_sample - 1))).astype(np.int16)
kaldi_helper = ivector_PLDA_kaldiHelper(pre_model_dir=self.pre_model_dir, audio_dir=self.audio_dir, mfcc_dir=self.mfcc_dir, log_dir=self.log_dir, ivector_dir=self.ivector_dir)
score_array = kaldi_helper.score(audio_list, self.utt_ids, n_jobs=n_jobs, flag=1, train_ivector_scp=self.train_ivector_scp, debug=debug)
score_array = (score_array - self.z_norm_means) / self.z_norm_stds
return score_array # (n_audios, n_spks) or (n_spks, )
path = os.path.join(z_norm_dir, audio_name)
z_norm_utt_path.append(path)
z_norm_utt_id.append(utt_id)
z_norm_spk_id.append(spk_id)
audio_path_list = (enroll_utt_path + z_norm_utt_path)
spk_id_list = (enroll_spk_id + z_norm_spk_id)
utt_id_list = (enroll_utt_id + z_norm_utt_id)
''' step 1: generate ivector identity (stored in ivector_dir) and corresponding speaker model (stored as model/XX.iv)
'''
print(
"----- step 1: generate ivector identity and corresponding speaker model -----"
)
iv_helper = ivector_PLDA_kaldiHelper(audio_dir=audio_dir,
mfcc_dir=mfcc_dir,
log_dir=log_dir,
ivector_dir=ivector_dir)
print("--- extracting and scoring ---")
iv_helper.score_existing(audio_path_list,
enroll_utt_id,
spk_id_list=spk_id_list,
utt_id_list=utt_id_list,
test_utt_id=z_norm_utt_id,
n_jobs=n_jobs,
flag=1,
debug=debug)
print("--- extracting and scoring done---")
print("--- resolve score and obtain z norm mean and std value ---")
