def test_preprocessing():
# make arguments
args = make_args()
# prepare dummy wav files
wavdir = "data/wav"
if not os.path.exists(wavdir):
os.makedirs(wavdir)
for i in range(5):
make_dummy_wav(wavdir + "/%d.wav" % i, 8000, args.fs)
# feature extract
wav_list = find_files(wavdir, "*.wav")
if not os.path.exists(args.wavdir):
os.makedirs(args.wavdir)
world_feature_extract(wav_list, args)
# calc_stats
file_list = find_files(args.hdf5dir, "*.h5")
calc_stats(file_list, args)
# noise shaping
wav_list = find_files(args.wavdir, "*.wav")
if not os.path.exists(args.writedir):
os.makedirs(args.writedir)
noise_shaping(wav_list, args)
# assert list length
wav_ns_list = find_files(args.writedir, "*.wav")
assert len(wav_list) == len(file_list)
assert len(wav_list) == len(wav_ns_list)
def test_preprocessing():
# make arguments
args = make_args()
# prepare dummy wav files
wavdir = "data/wav"
if not os.path.exists(wavdir):
os.makedirs(wavdir)
for i in range(5):
make_dummy_wav(wavdir + "/%d.wav" % i, 8000, args.fs)
# feature extract
wav_list = find_files(wavdir, "*.wav")
if not os.path.exists(args.wavdir):
os.makedirs(args.wavdir)
args.feature_type = "world"
world_feature_extract(wav_list, args)
args.feature_type = "melspc"
melspectrogram_extract(wav_list, args)
args.feature_type = "mcep"
melcepstrum_extract(wav_list, args)
# calc_stats
file_list = find_files(args.hdf5dir, "*.h5")
args.feature_type = "world"
calc_stats(file_list, args)
args.feature_type = "melspc"
calc_stats(file_list, args)
args.feature_type = "mcep"
calc_stats(file_list, args)
# noise shaping
wav_list = find_files(args.wavdir, "*.wav")
args.feature_type = "world"
args.writedir = "data/wav_ns/world"
if not os.path.exists(args.writedir):
os.makedirs(args.writedir)
world_noise_shaping(wav_list, args)
args.feature_type = "mcep"
args.writedir = "data/wav_ns/mcep"
if not os.path.exists(args.writedir):
os.makedirs(args.writedir)
melcepstrum_noise_shaping(wav_list, args)
def test_train_generator():
# make dummy wavfiles
wavdir = "data/wav"
if not os.path.exists(wavdir):
os.makedirs(wavdir)
for i in range(5):
make_dummy_wav(wavdir + "/%d.wav" % i)
# make features
feat_args = make_feature_args()
wav_list = find_files(wavdir, "*.wav")
if not os.path.exists(feat_args.wavdir):
os.makedirs(feat_args.wavdir)
feat_args.feature_type = "melspc"
melspectrogram_extract(wav_list, feat_args)
feat_args.feature_type = "world"
world_feature_extract(wav_list, feat_args)
feat_list = find_files(feat_args.hdf5dir, "*.h5")
for ft in ["world", "melspc"]:
# ----------------------------------
# minibatch without upsampling layer
# ----------------------------------
generator_args = make_train_generator_args(wav_list=wav_list,
feat_list=feat_list,
feature_type=ft,
use_upsampling_layer=False,
batch_length=10000,
batch_size=5)
generator = train_generator(**vars(generator_args))
(x, h), t = next(generator)
assert x.size(0) == t.size(0) == h.size(0)
assert x.size(1) == t.size(1) == h.size(2)
# ----------------------------------------
# utterance batch without upsampling layer
# ----------------------------------------
generator_args = make_train_generator_args(wav_list=wav_list,
feat_list=feat_list,
feature_type=ft,
use_upsampling_layer=False,
batch_length=None,
batch_size=5)
generator = train_generator(**vars(generator_args))
(x, h), t = next(generator)
assert x.size(0) == t.size(0) == h.size(0) == 1
assert x.size(1) == t.size(1) == h.size(2)
# -------------------------------
# minibatch with upsampling layer
# -------------------------------
generator_args = make_train_generator_args(wav_list=wav_list,
feat_list=feat_list,
feature_type=ft,
use_upsampling_layer=True,
batch_length=10000,
batch_size=5)
generator = train_generator(**vars(generator_args))
(x, h), t = next(generator)
assert x.size(0) == t.size(0) == h.size(0)
assert x.size(1) == t.size(
1) == h.size(2) * generator_args.upsampling_factor
# -------------------------------------
# utterance batch with upsampling layer
# -------------------------------------
generator_args = make_train_generator_args(wav_list=wav_list,
feat_list=feat_list,
feature_type=ft,
use_upsampling_layer=True,
batch_length=None,
batch_size=5)
generator = train_generator(**vars(generator_args))
(x, h), t = next(generator)
assert x.size(0) == t.size(0) == h.size(0) == 1
assert x.size(1) == t.size(
1) == h.size(2) * generator_args.upsampling_factor
def test_decode_generator():
# make dummy wavfiles
wavdir = "data/wav"
if not os.path.exists(wavdir):
os.makedirs(wavdir)
for i in range(5):
make_dummy_wav(wavdir + "/%d.wav" % i)
# make features
feat_args = make_feature_args()
wav_list = find_files(wavdir, "*.wav")
if not os.path.exists(feat_args.wavdir):
os.makedirs(feat_args.wavdir)
feat_args.feature_type = "melspc"
melspectrogram_extract(wav_list, feat_args)
feat_args.feature_type = "world"
world_feature_extract(wav_list, feat_args)
feat_list = find_files(feat_args.hdf5dir, "*.h5")
for ft in ["world", "melspc"]:
# ----------------------------------
# non-batch without upsampling layer
# ----------------------------------
generator_args = make_decode_generator_args(feat_list=feat_list,
feature_type=ft,
use_upsampling_layer=False,
batch_size=1)
generator = decode_generator(**vars(generator_args))
_, (x, h, n_samples) = next(generator)
assert x.size(0) == h.size(0) == 1
assert h.size(2) == n_samples + 1
# -------------------------------
# non-batch with upsampling layer
# -------------------------------
generator_args = make_decode_generator_args(feat_list=feat_list,
feature_type=ft,
use_upsampling_layer=True,
batch_size=1)
generator = decode_generator(**vars(generator_args))
_, (x, h, n_samples) = next(generator)
assert x.size(0) == h.size(0) == 1
assert h.size(2) * generator_args.upsampling_factor == n_samples + 1
# ----------------------------------
# minibatch without upsampling layer
# ----------------------------------
generator_args = make_decode_generator_args(feat_list=feat_list,
feature_type=ft,
use_upsampling_layer=False,
batch_size=5)
generator = decode_generator(**vars(generator_args))
_, (batch_x, batch_h, n_samples_list) = next(generator)
assert batch_x.size(0) == batch_h.size(0) == len(n_samples_list)
assert batch_h.size(2) == max(n_samples_list) + 1
# -------------------------------
# minibatch with upsampling layer
# -------------------------------
generator_args = make_decode_generator_args(feat_list=feat_list,
feature_type=ft,
use_upsampling_layer=True,
batch_size=5)
generator = decode_generator(**vars(generator_args))
_, (batch_x, batch_h, n_samples_list) = next(generator)
assert batch_x.size(0) == batch_h.size(0) == len(n_samples_list)
assert batch_h.size(2) * generator_args.upsampling_factor == max(
n_samples_list) + 1