import os
import sys
import numpy as np
from scipy.io.wavfile import write as wavwrite
import tensorflow as tf
out_dir, tfrecord_fps = sys.argv[1], sys.argv[2:]
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
def _mapper(example_proto):
features = {
'samples': tf.FixedLenSequenceFeature([1], tf.float32, allow_missing=True),
'label': tf.FixedLenSequenceFeature([], tf.string, allow_missing=True)
}
example = tf.parse_single_example(example_proto, features)
wav = example['samples'][:, 0]
wav = wav[:16384]
wav_len = tf.shape(wav)[0]
wav = tf.pad(wav, [[0, 16384 - wav_len]])
label = tf.reduce_join(example['label'], 0)
return wav, label
dataset = tf.data.TFRecordDataset(tfrecord_fps)
dataset = dataset.map(_mapper)
dataset = dataset.apply(tf.contrib.data.batch_and_drop_remainder(1))
x, y = dataset.make_one_shot_iterator().get_next()
x, y = x[0], y[0]
with tf.Session() as sess:
i = 0
while True:
try:
_x, _y = sess.run([x, y])
except:
break
_x *= 32767.
_x = np.clip(_x, -32767., 32767.)
_x = _x.astype(np.int16)
wavwrite(os.path.join(out_dir, '{}_{}.wav'.format(_y, str(i).zfill(5))), 16000, _x)
i += 1
