def main():
parser = argparse.ArgumentParser()
parser.add_argument('--fs', type=int, default=22050,
help='Sampling frequency')
parser.add_argument('--fmax', type=int, default=None, nargs='?',
help='Maximum frequency')
parser.add_argument('--fmin', type=int, default=None, nargs='?',
help='Minimum frequency')
parser.add_argument('--n_fft', type=int, default=1024,
help='FFT length in point')
parser.add_argument('--n_shift', type=int, default=512,
help='Shift length in point')
parser.add_argument('--win_length', type=int, default=None, nargs='?',
help='Analisys window length in point')
parser.add_argument('--n_mels', type=int, default=80,
help='Number of mel basis')
parser.add_argument('--window', type=str, default='hann',
choices=['hann', 'hamming'],
help='Type of window')
parser.add_argument('scp', type=str,
help='Feat scp files')
parser.add_argument('outdir', type=str,
help='Output directory')
args = parser.parse_args()
# logging info
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s")
# load scp
reader = kaldi_io_py.read_mat_scp(args.scp)
# chech direcitory
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
# extract feature and then write as ark with scp format
for idx, (utt_id, lmspc) in enumerate(reader, 1):
spc = logmelspc_to_linearspc(
lmspc,
fs=args.fs,
n_mels=args.n_mels,
n_fft=args.n_fft,
fmin=args.fmin,
fmax=args.fmax)
y = griffin_lim(
spc,
n_fft=args.n_fft,
n_shift=args.n_shift,
win_length=args.win_length,
window=args.window)
logging.info("(%d) %s" % (idx, utt_id))
write(args.outdir + "/%s.wav" % utt_id,
args.fs,
(y * np.iinfo(np.int16).max).astype(np.int16))
def test_voxforge_feats():
import kaldi_io_py
pytest.importorskip("kaldi_io")
import kaldi_io
train_scp = "scp:egs/voxforge/asr1/data/tr_it/feats.scp"
if not os.path.exists(train_scp):
pytest.skip("voxforge scp has not been created")
r1 = kaldi_io_py.read_mat_scp(train_scp)
r2 = kaldi_io.RandomAccessBaseFloatMatrixReader(train_scp)
for k, v1 in r1:
k = str(k)
print(k)
v2 = r2[k]
assert v1.shape == v2.shape
numpy.testing.assert_allclose(v1, v2, atol=1e-5)
# coding: utf-8
import numpy
import kaldi_io
import kaldi_io_py
import lazy_io
train_scp = "scp:egs/voxforge/asr1/data/tr_it/feats.scp"
r1 = kaldi_io_py.read_mat_scp(train_scp)
r2 = kaldi_io.RandomAccessBaseFloatMatrixReader(train_scp)
r3 = lazy_io.read_dict_scp(train_scp)
for k, v1 in r1:
k = str(k)
print(k)
v2 = r2[k]
v3 = r3[k]
assert v1.shape == v2.shape
assert v1.shape == v3.shape
numpy.testing.assert_allclose(v1, v2, atol=1e-5)
numpy.testing.assert_allclose(v1, v3, atol=0)
type=str,
required=True,
help='output file')
parser.add_argument('--cmvn', type=str, help='cmvn file')
args = parser.parse_args()
#ark='ark: copy-feats --print-args=false scp:%s ark:- | ' % (args.input)
#print(args.input)
# read global mean/stddev for normalize
if args.cmvn:
with h5py.File(args.cmvn, 'r') as f:
cmvn_mean = f['mean'].value
cmvn_dev = f['stddev'].value
generator = kaldi_io_py.read_mat_scp(args.input)
with h5py.File(args.output, 'w') as f:
for key, mat in generator:
if args.cmvn:
cp = np.ndarray(mat.shape)
for row in range(mat.shape[0]):
cp[row] = mat[row] - cmvn_mean.reshape(mat.shape[1], )
cp[row] = mat[row] / cmvn_dev.reshape(mat.shape[1], )
f.create_dataset(key,
data=cp,
compression='gzip',
compression_opts=9)
else:
f.create_dataset(key,
data=mat,
compression='gzip',
spk_file=os.path.join(args.outdir,'speakers')
with open(spk_file, 'w') as wf:
with open(args.feats, 'r') as f:
lines=f.readlines()
for line in lines:
spk = re.sub('\S\S\S$', "", re.sub('_DT',"",line.split()[0]))
if spk in speakers:
speakers[spk]=spk_num
out="{0} {1}\n".format(spk, spk_num)
wf.write(out)
spk_num+=1
hdf_file=os.path.join(args.outdir,'data.h5')
generator=kaldi_io_py.read_mat_scp(args.feats)
with h5py.File(hdf_file, 'w') as hdf:
for key, mat in generator:
label = self.textset[key]
hdf.create_group(key)
hdf.create_dataset(key+'/data', data=mat, compression='gzip', compression_opts=9)
if '_DT' in key:
label=0
else:
label=1 # 1 if speaker is deaf
spk=re.sub('\S\S\S$', "", re.sub('_DT',"", key))
spk_label=speakers[$spk]
hdf.create_dataset(key+'/label', data=label)
hdf.create_dataset(key+'/speaker',data=spk_label)
