def main(*argv):
argv = argv if argv else sys.argv[1:]
# Options for python
description = 'Train GMM and converted GV statistics'
parser = argparse.ArgumentParser(description=description)
parser.add_argument('org_list_file',
type=str,
help='List file of original speaker')
parser.add_argument('pair_yml',
type=str,
help='Yml file of the speaker pair')
parser.add_argument('pair_dir',
type=str,
help='Directory path of h5 files')
args = parser.parse_args(argv)
# read pair-dependent yml file
pconf = PairYML(args.pair_yml)
# read joint feature vector
jntf = os.path.join(args.pair_dir, 'jnt',
'it' + str(pconf.jnt_n_iter + 1) + '_jnt.h5')
jnth5 = HDF5(jntf, mode='r')
jnt = jnth5.read(ext='jnt')
# train GMM for mcep using joint feature vector
gmm = GMMTrainer(n_mix=pconf.GMM_mcep_n_mix,
n_iter=pconf.GMM_mcep_n_iter,
covtype=pconf.GMM_mcep_covtype)
gmm.train(jnt)
# save GMM
gmm_dir = os.path.join(args.pair_dir, 'model')
if not os.path.exists(gmm_dir):
os.makedirs(gmm_dir)
gmmpath = os.path.join(gmm_dir, 'GMM.pkl')
joblib.dump(gmm.param, gmmpath)
print("Conversion model save into " + gmmpath)
# calculate GV statistics of converted feature
h5_dir = os.path.join(args.pair_dir, 'h5')
org_mceps = read_feats(args.org_list_file, h5_dir, ext='mcep')
cv_mceps = feature_conversion(pconf, org_mceps, gmm, gmmmode=None)
diffcv_mceps = feature_conversion(pconf, org_mceps, gmm, gmmmode='diff')
gv = GV()
cvgvstats = gv.estimate(cv_mceps)
diffcvgvstats = gv.estimate(diffcv_mceps)
# open h5 files
statspath = os.path.join(gmm_dir, 'cvgv.h5')
cvgvh5 = HDF5(statspath, mode='w')
cvgvh5.save(cvgvstats, ext='cvgv')
cvgvh5.save(diffcvgvstats, ext='diffcvgv')
print("Converted gvstats save into " + statspath)
def main(*argv):
argv = argv if argv else sys.argv[1:]
# Options for python
description = 'estimate joint feature of source and target speakers'
parser = argparse.ArgumentParser(description=description)
parser.add_argument('pair_yml',
type=str,
help='Yml file of the speaker pair')
parser.add_argument('org_list_file',
type=str,
help='List file of original speaker')
parser.add_argument('tar_list_file',
type=str,
help='List file of target speaker')
parser.add_argument('pair_dir',
type=str,
help='Directory path of h5 files')
args = parser.parse_args(argv)
# read pair-dependent yml file
pconf = PairYML(args.pair_yml)
# read source and target features from HDF file
h5_dir = os.path.join(args.pair_dir, 'h5')
org_mceps = read_feats(args.org_list_file, h5_dir, ext='mcep')
org_npows = read_feats(args.org_list_file, h5_dir, ext='npow')
tar_mceps = read_feats(args.tar_list_file, h5_dir, ext='mcep')
tar_npows = read_feats(args.tar_list_file, h5_dir, ext='npow')
assert len(org_mceps) == len(tar_mceps)
assert len(org_npows) == len(tar_npows)
assert len(org_mceps) == len(org_npows)
itnum = 1
sd = 1 # start dimension for aligment of mcep
num_files = len(org_mceps)
print('{}-th joint feature extraction starts.'.format(itnum))
# first iteration
for i in range(num_files):
jdata, _, mcd = get_aligned_jointdata(org_mceps[i][:,
sd:], org_npows[i],
tar_mceps[i][:,
sd:], tar_npows[i])
print('distortion [dB] for {}-th file: {}'.format(i + 1, mcd))
if i == 0:
jnt = jdata
else:
jnt = np.r_[jnt, jdata]
itnum += 1
# second through final iteration
while itnum < pconf.jnt_n_iter + 1:
print('{}-th joint feature extraction starts.'.format(itnum))
# train GMM
trgmm = GMMTrainer(n_mix=pconf.GMM_mcep_n_mix,
n_iter=pconf.GMM_mcep_n_iter,
covtype=pconf.GMM_mcep_covtype)
trgmm.train(jnt)
cvgmm = GMMConvertor(n_mix=pconf.GMM_mcep_n_mix,
covtype=pconf.GMM_mcep_covtype)
cvgmm.open_from_param(trgmm.param)
twfs = []
for i in range(num_files):
cvmcep = cvgmm.convert(static_delta(org_mceps[i][:, sd:]),
cvtype=pconf.GMM_mcep_cvtype)
jdata, twf, mcd = get_aligned_jointdata(org_mceps[i][:, sd:],
org_npows[i],
tar_mceps[i][:, sd:],
tar_npows[i],
cvdata=cvmcep)
print('distortion [dB] for {}-th file: {}'.format(i + 1, mcd))
if i == 0:
jnt = jdata
else:
jnt = np.r_[jnt, jdata]
twfs.append(twf)
itnum += 1
# save joint feature vector
jnt_dir = os.path.join(args.pair_dir, 'jnt')
if not os.path.exists(jnt_dir):
os.makedirs(jnt_dir)
jntpath = os.path.join(jnt_dir, 'it' + str(itnum) + '_jnt.h5')
jnth5 = HDF5(jntpath, mode='w')
jnth5.save(jnt, ext='jnt')
jnth5.close()
# save GMM
gmm_dir = os.path.join(args.pair_dir, 'GMM')
if not os.path.exists(gmm_dir):
os.makedirs(gmm_dir)
gmmpath = os.path.join(gmm_dir, 'it' + str(itnum) + '_gmm.pkl')
joblib.dump(trgmm.param, gmmpath)
# save twf
twf_dir = os.path.join(args.pair_dir, 'twf')
if not os.path.exists(twf_dir):
os.makedirs(twf_dir)
with open(args.org_list_file, 'r') as fp:
for line, twf in zip(fp, twfs):
f = os.path.basename(line.rstrip())
twfpath = os.path.join(twf_dir,
'it' + str(itnum) + '_' + f + '.h5')
twfh5 = HDF5(twfpath, mode='w')
twfh5.save(twf, ext='twf')
twfh5.close()
def main(*argv):
argv = argv if argv else sys.argv[1:]
# Options for python
description = 'estimate joint feature of source and target speakers'
parser = argparse.ArgumentParser(description=description)
parser.add_argument('org_yml', type=str,
help='Yml file of the original speaker')
parser.add_argument('tar_yml', type=str,
help='Yml file of the target speaker')
parser.add_argument('pair_yml', type=str,
help='Yml file of the speaker pair')
parser.add_argument('org_list_file', type=str,
help='List file of original speaker')
parser.add_argument('tar_list_file', type=str,
help='List file of target speaker')
parser.add_argument('pair_dir', type=str,
help='Directory path of h5 files')
args = parser.parse_args(argv)
# read speaker-dependent yml files
oconf = SpeakerYML(args.org_yml)
tconf = SpeakerYML(args.tar_yml)
# read pair-dependent yml file
pconf = PairYML(args.pair_yml)
# read source and target features from HDF file
h5_dir = os.path.join(args.pair_dir, 'h5')
org_mceps = read_feats(args.org_list_file, h5_dir, ext='mcep')
org_npows = read_feats(args.org_list_file, h5_dir, ext='npow')
tar_mceps = read_feats(args.tar_list_file, h5_dir, ext='mcep')
tar_npows = read_feats(args.tar_list_file, h5_dir, ext='npow')
assert len(org_mceps) == len(tar_mceps)
assert len(org_npows) == len(tar_npows)
assert len(org_mceps) == len(org_npows)
# dtw between original and target w/o 0th and silence
print('## Alignment mcep w/o 0-th and silence ##')
jmceps, twfs = align_feature_vectors(org_mceps,
org_npows,
tar_mceps,
tar_npows,
pconf,
opow=oconf.power_threshold,
tpow=tconf.power_threshold,
itnum=pconf.jnt_n_iter,
sd=1,
)
jnt_mcep = transform_jnt(jmceps)
# create joint feature for codeap using given twfs
print('## Alignment codeap using given twf ##')
org_codeaps = read_feats(args.org_list_file, h5_dir, ext='codeap')
tar_codeaps = read_feats(args.tar_list_file, h5_dir, ext='codeap')
jcodeaps = []
for i in range(len(org_codeaps)):
# extract codeap joint feature vector
jcodeap, _, _ = get_alignment(org_codeaps[i],
org_npows[i],
tar_codeaps[i],
tar_npows[i],
opow=oconf.power_threshold,
tpow=tconf.power_threshold,
given_twf=twfs[i])
jcodeaps.append(jcodeap)
jnt_codeap = transform_jnt(jcodeaps)
# save joint feature vectors
jnt_dir = os.path.join(args.pair_dir, 'jnt')
os.makedirs(jnt_dir, exist_ok=True)
jntpath = os.path.join(jnt_dir, 'it' + str(pconf.jnt_n_iter) + '_jnt.h5')
jnth5 = HDF5(jntpath, mode='a')
jnth5.save(jnt_mcep, ext='mcep')
jnth5.save(jnt_codeap, ext='codeap')
jnth5.close()
# save twfs
twf_dir = os.path.join(args.pair_dir, 'twf')
os.makedirs(twf_dir, exist_ok=True)
with open(args.org_list_file, 'r') as fp:
for line, twf in zip(fp, twfs):
f = os.path.basename(line.rstrip())
twfpath = os.path.join(
twf_dir, 'it' + str(pconf.jnt_n_iter) + '_' + f + '.h5')
twfh5 = HDF5(twfpath, mode='a')
twfh5.save(twf, ext='twf')
twfh5.close()