def create_tfrecord(self, dataset_list, tfrecord_path):
print("Start converting...")
options = tf.python_io.\
TFRecordOptions(compression_type=tf.python_io.TFRecordCompressionType.GZIP)
writer = tf.python_io.TFRecordWriter(path=tfrecord_path, options=options)
def round_up(length, reduction_factor):
remain = length % reduction_factor
return reduction_factor - remain
for dataset in dataset_list:
audio_file_path = dataset["audio_file_path"]
_, mel, lin = ap.get_features(audio_file_path)
padding_length = round_up(np.shape(mel)[-1], hp.reduction_factor)
mel = np.pad(mel, ((0, 0), (0, padding_length)), "constant", constant_values=0.)
lin = np.pad(lin, ((0, 0), (0, padding_length)), "constant", constant_values=0.)
script = dataset["normalized_script"]
script = sp.encode_script(script, char2idx=self._char2idx)
script = np.asarray(script, dtype=np.int32)
example = tf.train.Example(
features=tf.train.Features(
feature={
"mel": _bytes_feature(mel.tostring()),
"lin": _bytes_feature(lin.tostring()),
"script": _bytes_feature(script.tostring())
}
)
)
writer.write(example.SerializeToString())
writer.close()
print("Done...")
def _generate_batch(self):
def round_up(length, reduction_factor):
remain = length % reduction_factor
return reduction_factor - remain
while True:
random_dataset_list = self._dataset_list[:]
random.shuffle(random_dataset_list)
for dataset in random_dataset_list:
audio_file_path = dataset["audio_file_path"]
_, mel, lin = ap.get_features(audio_file_path)
padding_length = round_up(
np.shape(mel)[-1], hp.reduction_factor)
mel = np.pad(mel, ((0, 0), (0, padding_length)),
"constant",
constant_values=0.)
lin = np.pad(lin, ((0, 0), (0, padding_length)),
"constant",
constant_values=0.)
spk = np.asarray(self._spk2idx[dataset["spk"]], dtype=np.int32)
script = dataset["script"]
script = sp.encode_script(script, char2idx=self._char2idx)
script = np.asarray(script, dtype=np.int32)
yield mel, lin, spk, script
def main():
dataset_list = kss.get_dataset_list(hp.dataset_path)
idx2char, char2idx, idx2ord = kss.get_idx2char(dataset_list)
print(len(dataset_list))
print(len(idx2char))
print(idx2char)
#print([(char.encode("utf-8"), char) for char in idx2char])
#print([ord(char) for char in idx2char])
#print([elem for elem in zip(idx2char, idx2ord)])
random_dataset_list = dataset_list[:]
random.shuffle(random_dataset_list)
print([dataset["script"] for dataset in random_dataset_list[:5]])
print([dataset["normalized_script"] for dataset in random_dataset_list[:5]])
mels = []
lins = []
audio, mel, lin = ap.get_features(dataset_list[0]["audio_file_path"])
audio_inverse = ap.lin_to_audio(lin)
padding_length = round_up(np.shape(mel)[-1], hp.reduction_factor)
mel = np.pad(mel, ((0, 0), (0, padding_length)), "constant", constant_values=0.)
lin = np.pad(lin, ((0, 0), (0, padding_length)), "constant", constant_values=0.)
print(dataset_list[0]["audio_file_path"])
librosa.output.write_wav(
"%s/audio_sample.wav" % hp.audio_save_path, audio, hp.sampling_rate
)
librosa.output.write_wav(
"%s/audio_inverse_sample.wav" % hp.audio_save_path, audio_inverse, hp.sampling_rate
)
fig = plt.figure()
librosa.display.specshow(lin,
y_axis="linear",
x_axis="time",
sr=hp.sampling_rate,
hop_length=hp.hop_length)
plt.colorbar()
plt.tight_layout()
fig.savefig("%s/lin_sample.png" % hp.logdir_root)
plt.clf()
plt.cla()
plt.close()
fig = plt.figure()
librosa.display.specshow(mel,
y_axis="linear",
x_axis="time",
sr=hp.sampling_rate,
hop_length=hp.hop_length)
plt.colorbar()
plt.tight_layout()
fig.savefig("%s/mel_sample.png" % hp.logdir_root)
plt.clf()
plt.cla()
plt.close()
random.shuffle(dataset_list)
for dataset in dataset_list[:10]:
_, mel, lin = ap.get_features(dataset["audio_file_path"])
padding_length = round_up(np.shape(mel)[-1], hp.reduction_factor)
#mel = np.pad(mel, ((0, 0), (0, padding_length)), "constant", constant_values=0.)
#lin = np.pad(lin, ((0, 0), (0, padding_length)), "constant", constant_values=0.)
mel = np.reshape(mel, [-1]).tolist()
lin = np.reshape(lin, [-1]).tolist()
mels = mels + mel
lins = lins + lin
fig = plt.figure()
n , bins, patches = plt.hist(mels, 50, density=1, facecolor="blue", alpha=0.75)
fig.savefig("%s/hist_mel.png" % hp.logdir_root)
fig = plt.figure()
n , bins, patches = plt.hist(lins, 50, density=1, facecolor="blue", alpha=0.75)
fig.savefig("%s/hist_lin.png" % hp.logdir_root)