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 read_feats(listf, h5dir, ext='mcep'):
"""HDF5 handler
Create list consisting of arrays listed in the list
Parameters
---------
listf : str,
Path of list file
h5dir : str,
Path of hdf5 directory
ext : str,
`mcep` : mel-cepstrum
`f0` : F0
Returns
---------
datalist : list of arrays
"""
datalist = []
with open(listf, 'r') as fp:
for line in fp:
f = line.rstrip()
h5f = os.path.join(h5dir, f + '.h5')
h5 = HDF5(h5f, mode='r')
datalist.append(h5.read(ext))
h5.close()
return datalist
def _save_hdf5(self, h5f):
h5 = HDF5(h5f, "a")
for k, v in self.feats.items():
if v.dtype == np.float64:
v = v.astype(np.float32)
h5.save(v, ext=k)
h5.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('speaker', type=str, help='Input speaker label')
parser.add_argument('list_file',
type=str,
help='List file of the input speaker')
parser.add_argument('pair_dir',
type=str,
help='Statistics directory of the speaker')
args = parser.parse_args(argv)
# open h5 files
h5_dir = os.path.join(args.pair_dir, 'h5')
statspath = os.path.join(args.pair_dir, 'stats', args.speaker + '.h5')
h5 = HDF5(statspath, mode='a')
# estimate and save F0 statistics
f0stats = F0statistics()
f0s = read_feats(args.list_file, h5_dir, ext='f0')
f0stats = f0stats.estimate(f0s)
h5.save(f0stats, ext='f0stats')
print("f0stats save into " + statspath)
# estimate and save GV of orginal and target speakers
gv = GV()
mceps = read_feats(args.list_file, h5_dir, ext='mcep')
gvstats = gv.estimate(mceps)
h5.save(gvstats, ext='gv')
print("gvstats save into " + statspath)
h5.close()
def read_ppg_feats(s_listf, t_listf, h5dir, ext='mcep'):
s_datalist = []
t_datalist = []
with open(s_listf, 'r') as fp:
for line in fp:
s_mcep = []
t_mcep = []
f = line.rstrip()
h5f = os.path.join(h5dir, f + '.h5')
h5 = HDF5(h5f, mode='r')
s_mcep.append(h5.read(ext))
h5.close()
source = "BDL_1.27"
target = "RRBI_16k"
line = line.replace(source, target)
f = line.rstrip()
h5f = os.path.join(h5dir, f + '.h5')
h5 = HDF5(h5f, mode='r')
t_mcep.append(h5.read(ext))
h5.close()
source_post = "/home/anurag/kaldi/egs/librispeech/s5/post_source/"
target_post = "/home/anurag/kaldi/egs/librispeech/s5/post_target/"
f = "post." + f[-12:] + ".ark"
print(f)
s_post = os.path.join(source_post, f)
t_post = os.path.join(target_post, f)
s_post = np.loadtxt(s_post)
t_post = np.loadtxt(t_post)
for i in range(len(s_post)):
score = []
for j in range(len(t_post)):
score.append(KL(s_post[i], t_post[j]))
t_mcep.append(t_post[score.index(min(score))])
s_datalist.append(s_mcep)
t_datalist.append(t_mcep)
print(len(s_datalist))
print(len(t_datalist))
return s_mcep, t_mcep
def main(*argv):
argv = argv if argv else sys.argv
# Options for python
dcp = 'Extract aoucstic features for the speaker'
parser = argparse.ArgumentParser(description=dcp)
parser.add_argument('--overwrite',
default=False,
action='store_true',
help='Overwrite h5 file')
parser.add_argument('speaker', type=str, help='Input speaker label')
parser.add_argument('ymlf', type=str, help='Yml file of the input speaker')
parser.add_argument('list_file',
type=str,
help='List file of the input speaker')
parser.add_argument('wav_dir',
type=str,
help='Wav file directory of the speaker')
parser.add_argument('pair_dir',
type=str,
help='Directory of the speaker pair')
args = parser.parse_args(argv)
# read parameters from speaker yml
sconf = SpeakerYML(args.ymlf)
h5_dir = os.path.join(args.pair_dir, 'h5')
anasyn_dir = os.path.join(args.pair_dir, 'anasyn')
if not os.path.exists(os.path.join(h5_dir, args.speaker)):
os.makedirs(os.path.join(h5_dir, args.speaker))
if not os.path.exists(os.path.join(anasyn_dir, args.speaker)):
os.makedirs(os.path.join(anasyn_dir, args.speaker))
# constract FeatureExtractor class
feat = FeatureExtractor(analyzer=sconf.analyzer,
fs=sconf.wav_fs,
fftl=sconf.wav_fftl,
shiftms=sconf.wav_shiftms,
minf0=sconf.f0_minf0,
maxf0=sconf.f0_maxf0)
# constract Synthesizer class
synthesizer = Synthesizer(fs=sconf.wav_fs,
fftl=sconf.wav_fftl,
shiftms=sconf.wav_shiftms)
# open list file
with open(args.list_file, 'r') as fp:
for line in fp:
f = line.rstrip()
h5f = os.path.join(h5_dir, f + '.h5')
if (not os.path.exists(h5f)) or args.overwrite:
wavf = os.path.join(args.wav_dir, f + '.wav')
fs, x = wavfile.read(wavf)
x = np.array(x, dtype=np.float)
x = low_cut_filter(x, fs, cutoff=70)
assert fs == sconf.wav_fs
print("Extract acoustic features: " + wavf)
# analyze F0, spc, and ap
f0, spc, ap = feat.analyze(x)
mcep = feat.mcep(dim=sconf.mcep_dim, alpha=sconf.mcep_alpha)
npow = feat.npow()
codeap = feat.codeap()
# save features into a hdf5 file
h5 = HDF5(h5f, mode='w')
h5.save(f0, ext='f0')
# h5.save(spc, ext='spc')
# h5.save(ap, ext='ap')
h5.save(mcep, ext='mcep')
h5.save(npow, ext='npow')
h5.save(codeap, ext='codeap')
h5.close()
# analysis/synthesis using F0, mcep, and ap
wav = synthesizer.synthesis(
f0,
mcep,
ap,
alpha=sconf.mcep_alpha,
)
wav = np.clip(wav, -32768, 32767)
anasynf = os.path.join(anasyn_dir, f + '.wav')
wavfile.write(anasynf, fs, np.array(wav, dtype=np.int16))
else:
print("Acoustic features already exist: " + h5f)
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('-gmmmode',
'--gmmmode',
type=str,
default=None,
help='mode of the GMM [None, diff, or intra]')
parser.add_argument('org', type=str, help='Original speaker')
parser.add_argument('tar', type=str, help='Target speaker')
parser.add_argument('org_yml',
type=str,
help='Yml file of the original speaker')
parser.add_argument('pair_yml',
type=str,
help='Yml file of the speaker pair')
parser.add_argument('eval_list_file',
type=str,
help='List file for evaluation')
parser.add_argument('wav_dir',
type=str,
help='Directory path of source spekaer')
parser.add_argument('pair_dir',
type=str,
help='Directory path of pair directory')
args = parser.parse_args(argv)
# read parameters from speaker yml
sconf = SpeakerYML(args.org_yml)
pconf = PairYML(args.pair_yml)
# read GMM for mcep
mcepgmmpath = os.path.join(args.pair_dir, 'model/GMM_mcep.pkl')
mcepgmm = GMMConvertor(
n_mix=pconf.GMM_mcep_n_mix,
covtype=pconf.GMM_mcep_covtype,
gmmmode=args.gmmmode,
)
param = joblib.load(mcepgmmpath)
mcepgmm.open_from_param(param)
print("GMM for mcep conversion mode: {}".format(args.gmmmode))
# read F0 statistics
stats_dir = os.path.join(args.pair_dir, 'stats')
orgstatspath = os.path.join(stats_dir, args.org + '.h5')
orgstats_h5 = HDF5(orgstatspath, mode='r')
orgf0stats = orgstats_h5.read(ext='f0stats')
orgstats_h5.close()
# read F0 and GV statistics for target
tarstatspath = os.path.join(stats_dir, args.tar + '.h5')
tarstats_h5 = HDF5(tarstatspath, mode='r')
tarf0stats = tarstats_h5.read(ext='f0stats')
targvstats = tarstats_h5.read(ext='gv')
tarstats_h5.close()
# read GV statistics for converted mcep
cvgvstatspath = os.path.join(args.pair_dir, 'model', 'cvgv.h5')
cvgvstats_h5 = HDF5(cvgvstatspath, mode='r')
cvgvstats = cvgvstats_h5.read(ext='cvgv')
diffcvgvstats = cvgvstats_h5.read(ext='diffcvgv')
cvgvstats_h5.close()
mcepgv = GV()
f0stats = F0statistics()
# constract FeatureExtractor class
feat = FeatureExtractor(analyzer=sconf.analyzer,
fs=sconf.wav_fs,
fftl=sconf.wav_fftl,
shiftms=sconf.wav_shiftms,
minf0=sconf.f0_minf0,
maxf0=sconf.f0_maxf0)
# constract Synthesizer class
synthesizer = Synthesizer(fs=sconf.wav_fs,
fftl=sconf.wav_fftl,
shiftms=sconf.wav_shiftms)
# test directory
test_dir = os.path.join(args.pair_dir, 'test')
os.makedirs(os.path.join(test_dir, args.org), exist_ok=True)
# conversion in each evaluation file
with open(args.eval_list_file, 'r') as fp:
for line in fp:
# open wav file
f = line.rstrip()
wavf = os.path.join(args.wav_dir, f + '.wav')
fs, x = wavfile.read(wavf)
x = x.astype(np.float)
x = low_cut_filter(x, fs, cutoff=70)
assert fs == sconf.wav_fs
# analyze F0, mcep, and ap
f0, spc, ap = feat.analyze(x)
mcep = feat.mcep(dim=sconf.mcep_dim, alpha=sconf.mcep_alpha)
mcep_0th = mcep[:, 0]
# convert F0
cvf0 = f0stats.convert(f0, orgf0stats, tarf0stats)
# convert mcep
cvmcep_wopow = mcepgmm.convert(static_delta(mcep[:, 1:]),
cvtype=pconf.GMM_mcep_cvtype)
cvmcep = np.c_[mcep_0th, cvmcep_wopow]
# synthesis VC w/ GV
if args.gmmmode is None:
cvmcep_wGV = mcepgv.postfilter(cvmcep,
targvstats,
cvgvstats=cvgvstats,
alpha=pconf.GV_morph_coeff,
startdim=1)
wav = synthesizer.synthesis(
cvf0,
cvmcep_wGV,
ap,
rmcep=mcep,
alpha=sconf.mcep_alpha,
)
wavpath = os.path.join(test_dir, f + '_VC.wav')
# synthesis DIFFVC w/ GV
if args.gmmmode == 'diff':
cvmcep[:, 0] = 0.0
cvmcep_wGV = mcepgv.postfilter(mcep + cvmcep,
targvstats,
cvgvstats=diffcvgvstats,
alpha=pconf.GV_morph_coeff,
startdim=1) - mcep
wav = synthesizer.synthesis_diff(
x,
cvmcep_wGV,
rmcep=mcep,
alpha=sconf.mcep_alpha,
)
wavpath = os.path.join(test_dir, f + '_DIFFVC.wav')
# write waveform
if not os.path.exists(os.path.join(test_dir, f)):
os.makedirs(os.path.join(test_dir, f))
wav = np.clip(wav, -32768, 32767)
wavfile.write(wavpath, fs, wav.astype(np.int16))
print(wavpath)
def feat2hdf5(mlfb, hdf5, ext="feats"):
tdir, name = Path(hdf5).parent, Path(hdf5).stem
h5f = tdir / (str(name) + ".h5")
h5 = HDF5(str(h5f), "a")
h5.save(mlfb, ext=ext)
h5.close()
def _save_hdf5(self, h5f):
h5 = HDF5(h5f, "a")
for k, v in self.feats.items():
h5.save(v, ext=k)
h5.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('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()
def main():
if args.file == 'con':
file = 'converted'
elif args.file == 'tar':
file = 'target'
else:
raise ValueError("The file is incorrect")
feat = FeatureExtractor(analyzer='world',
fs=22050,
fftl=1024,
shiftms=5,
minf0=args.minf0,
maxf0=args.maxf0)
# constract Synthesizer class
synthesizer = Synthesizer(fs=22050, fftl=1024, shiftms=5)
# open list file
with open('./list/' + file + '.list', 'r') as fp:
for line in fp:
f = line.rstrip()
h5f = os.path.join('./' + file + '/h5f/', f + '.h5')
if (not os.path.exists(h5f)):
wavf = os.path.join('./' + file + '/wav/', f + '.wav')
fs, x = wavfile.read(wavf)
x = np.array(x, dtype=np.float)
x = low_cut_filter(x, fs, cutoff=70)
print("Extract acoustic features: " + wavf)
# analyze F0, spc, and ap
f0, spc, ap = feat.analyze(x)
mcep = feat.mcep(dim=34, alpha=0.544)
npow = feat.npow()
codeap = feat.codeap()
# save features into a hdf5 file
h5 = HDF5(h5f, mode='w')
h5.save(f0, ext='f0')
# h5.save(spc, ext='spc')
# h5.save(ap, ext='ap')
h5.save(mcep, ext='mcep')
h5.save(npow, ext='npow')
h5.save(codeap, ext='codeap')
h5.close()
# analysis/synthesis using F0, mcep, and ap
wav = synthesizer.synthesis(
f0,
mcep,
ap,
alpha=0.544,
)
wav = np.clip(wav, -32768, 32767)
anasynf = os.path.join('./' + file + '/anasyn/', f + '.wav')
wavfile.write(anasynf, fs, np.array(wav, dtype=np.int16))
else:
print("Acoustic features already exist: " + h5f)
