def convert_image(self, image_path):
img = cv2.imread(image_path)
text = pytesseract.image_to_string(img, lang='eng', config='--psm 6')
image_path = os.path.normpath(image_path)
file_name = image_path.split(os.sep)[-2]
tag_name = image_path.split(os.sep)[-1].split(".")[0]
file_path = os.path.join(self._text_out_dir, file_name)
if not os.path.exists(file_path):
os.makedirs(file_path)
text_file_path = os.path.join(file_path, tag_name + ".txt")
print_error(text_file_path)
fd = open(text_file_path, "w")
fd.write("%s" % text)
return text_file_path
def __init__(self, hparams=None, data_iterator=None):
"""
https://arxiv.org/abs/1508.04306
:param hparams:
:param data_iterator:
"""
# ITextFeature.__init__(self)
ModelBase.__init__(self, hparams=hparams)
ShabdaWavPairFeature.__init__(self)
self._hparams = HParams(hparams, self.default_hparams())
print_error(self._hparams)
self.lstm_hidden_size = self._hparams.lstm_hidden_size
self.batch_size = self._hparams.batch_size
self.p_keep_ff = self._hparams.p_keep_ff
self.p_keep_rc = self._hparams.p_keep_rc
self.neff = self._hparams.neff
self.embd_dim_k = self._hparams.embd_dim_k
self.frames_per_sample = self._hparams.frames_per_sample
self.weights = None
self.biases = None
def _get_speech_features(wav_file_1, wav_file_2, sampling_rate, frame_size,
neff, amp_fac, min_amp, threshold, global_mean,
global_std, frames_per_sample):
"""
:param args:
:return:
"""
# (wav_file_1, wav_file_2, sampling_rate, frame_size, neff, amp_fac, min_amp,
# threshold, global_mean, global_std, frames_per_sample) = args
try:
# TODO: experimental though - is multi-processing required here? to reduce IO
# print_info("{} {}".format(wav_file_1, wav_file_2))
def _get_data():
speech_1 = _get_speech_data(wav_file_1, sampling_rate)
speech_2 = _get_speech_data(wav_file_2, sampling_rate)
# print(wav_file_1, wav_file_2)
# find the minimum length of two speeches
length = min(len(speech_1), len(speech_2))
#trim both the speeches to the minimum length
speech_1 = speech_1[:length]
speech_2 = speech_2[:length]
# mix the signals
speech_mix = speech_1 + speech_2
#get the spectral features in dB
speech_1_features = _get_log_spectrum_features(
speech_1, frame_size, neff, amp_fac, min_amp)
speech_2_features = _get_log_spectrum_features(
speech_2, frame_size, neff, amp_fac, min_amp)
speech_mix_features = _get_log_spectrum_features(
speech_mix, frame_size, neff, amp_fac, min_amp)
max_mag = np.max(speech_mix_features)
# apply threshold to the feature signal, to find the silent portion of the signal and
# construct a boolean array as a feature
# https://en.wikipedia.org/wiki/Voice_activity_detection
speech_voice_activity_detection = (speech_mix_features >
(max_mag - threshold))
# normalize the signal values with given global mean and std
speech_mix_features_final = (speech_mix_features -
global_mean) / global_std
number_frames = speech_1_features.shape[0]
new_data = _get_speech_samples(speech_1_features,
speech_2_features,
frames_per_sample, number_frames,
speech_mix_features_final,
speech_voice_activity_detection)
# print_error("deleting speech_1_features, speech_2_features, speech_voice_activity_detection, speech_mix_features")
speech_1_features = speech_2_features = speech_voice_activity_detection = speech_mix_features = speech_mix_features_final = None
return new_data
new_data = _get_data()
except (KeyboardInterrupt, SystemExit):
raise
except Exception as err:
print_error(traceback.print_exc())
new_data = []
return new_data
def generate_features(self, wav_file_1, wav_file_2):
try:
start = time.time()
speech_1, _ = librosa.core.load(wav_file_1,
sr=self._hparams.sampling_rate)
# amp factor between -3 dB - 3 dB
fac = np.random.rand(1)[0] * 6 - 3
speech_1 = 10.**(fac / 20) * speech_1
speech_2, _ = librosa.core.load(wav_file_2,
sr=self._hparams.sampling_rate)
fac = np.random.rand(1)[0] * 6 - 3
speech_2 = 10.**(fac / 20) * speech_2
# mix
length = min(len(speech_1), len(speech_2))
speech_1 = speech_1[:length]
speech_2 = speech_2[:length]
speech_mix = speech_1 + speech_2
# compute log spectrum for 1st speaker
speech_1_features = np.abs(
stft(speech_1,
self._hparams.frame_size)[:, :self._hparams.neff])
speech_1_features = np.maximum(
speech_1_features,
np.max(speech_1_features) / self._hparams.min_amp)
speech_1_features = 20. * np.log10(
speech_1_features * self._hparams.amp_fac)
# same for the 2nd speaker
speech_2_features = np.abs(
stft(speech_2,
self._hparams.frame_size)[:, :self._hparams.neff])
speech_2_features = np.maximum(
speech_2_features,
np.max(speech_2_features) / self._hparams.min_amp)
speech_2_features = 20. * np.log10(
speech_2_features * self._hparams.amp_fac)
# same for the mixture
speech_mix_spec0 = stft(
speech_mix, self._hparams.frame_size)[:, :self._hparams.neff]
speech_mix_features = np.abs(speech_mix_spec0)
# speech_phase = speech_mix_spec0 / speech_mix_spec
speech_mix_features = np.maximum(
speech_mix_features,
np.max(speech_mix_features) / self._hparams.min_amp)
speech_mix_features = 20. * np.log10(
speech_mix_features * self._hparams.amp_fac)
max_mag = np.max(speech_mix_features)
# if np.isnan(max_mag):
# import ipdb; ipdb.set_trace()
speech_VAD = (speech_mix_features >
(max_mag - self._hparams.threshold)).astype(int)
speech_mix_features = (
speech_mix_features -
self._hparams.global_mean) / self._hparams.global_std
#The ideal binary mask gives ownership of a time-frequency bin to the source whose magnitude is
# maximum among all sources in that bin.
# The mask values were assigned with 1 for active and 0 otherwise (binary),
# making Y x Y^T as the ideal affinity matrix for the mixture.
Y = np.array([
speech_1_features > speech_2_features,
speech_1_features < speech_2_features
]).astype('bool')
Y = np.transpose(Y, [1, 2, 0]).astype('bool')
# speech_mix_features = speech_mix_features[0:self._hparams.dummy_slicing_dim, :]
# speech_VAD = speech_VAD[0:self._hparams.dummy_slicing_dim, :]
# Y = Y[0:self._hparams.dummy_slicing_dim, :, :]
# print_info("{} vs {}".format(wav_file_1, wav_file_2))
end = time.time()
print_info("Thread name: {} : took {}".format(
threading.currentThread().getName(), end - start))
if speech_mix_features.shape[0] != 1247 or speech_VAD.shape[
0] != 1247 or Y.shape[0] != 1247:
raise Exception("Found files with improper duration/data")
return speech_mix_features.astype('float32'), speech_VAD.astype(
'bool'), Y.astype('bool')
except Exception as e:
print_warn(e)
print_error("{} vs {}".format(wav_file_1, wav_file_2))
return np.random.random((self._hparams.dummy_slicing_dim,129)).astype('float32'), \
np.empty((self._hparams.dummy_slicing_dim,129), dtype="bool"), \
np.empty((self._hparams.dummy_slicing_dim,129, 2), dtype="bool")
def discriminator(self, x, out_channel_dim, is_training=True, reuse=False):
# It must be Auto-Encoder style architecture
# Architecture : (64)4c2s-FC32_BR-FC64*14*14_BR-(1)4dc2s_S
with tf.variable_scope("namespace_discriminator", reuse=reuse):
# net = tf.nn.relu(conv2d(x, 64, 4, 4, 2, 2, name='d_conv1'))
net = tf.layers.conv2d(
x,
64,
4,
strides=2,
padding='same',
kernel_initializer=tf.random_normal_initializer(stddev=0.02),
name='d_conv1')
net = tf.nn.relu(net)
tf.logging.info("======> net: {}".format(net))
print_error("net1: {} ".format(net))
size = (self.image_size // 2)
net = tf.reshape(
net, [self._data_iterator.batch_size, size * size * 64])
# code = tf.nn.relu(bn(linear(net, 32, scope='d_fc6'), is_training=is_training, scope='d_bn6'))
code = tf.contrib.layers.fully_connected(inputs=net,
num_outputs=32,
scope="d_fc6")
code = tf.contrib.layers.batch_norm(code,
decay=0.9,
updates_collections=None,
epsilon=1e-5,
scale=True,
is_training=is_training,
scope='d_bn6')
code = tf.nn.relu(code)
print_error("code: {} ".format(code))
# net = tf.nn.relu(bn(linear(code, 64 * 14 * 14, scope='d_fc3'), is_training=is_training, scope='d_bn3'))
size = (self.image_size // 2)
net = tf.contrib.layers.fully_connected(inputs=code,
num_outputs=64 * size *
size,
scope="d_fc3")
net = tf.contrib.layers.batch_norm(net,
decay=0.9,
updates_collections=None,
epsilon=1e-5,
scale=True,
is_training=is_training,
scope='d_bn3')
print_error("net: {} ".format(net))
print_error(net)
size = (self.image_size // 2)
net = tf.reshape(net,
[self._data_iterator.batch_size, size, size, 64])
print_error(net)
# out = tf.nn.sigmoid(deconv2d(net, [self.gan_config.batch_size, 28, 28, 1], 4, 4, 2, 2, name='d_dc5'))
net = tf.layers.conv2d_transpose(net,
out_channel_dim,
4,
strides=2,
padding='same',
name='d_dc5')
out = tf.nn.sigmoid(net)
print_info("==================================")
print_info(out)
print_info(x)
# recon loss
recon_error = tf.sqrt(
2 * tf.nn.l2_loss(out - x)) / self._data_iterator.batch_size
print_info("==================================")
print_error(recon_error)
return out, recon_error, code
def visulaize(self, executor, file_path):
"""
:param executor:
:param test_file_path:
:return:
"""
estimator = executor.estimator
in_data_features, voice_activity_detection_data_features, phase_features = self._get_predict_samples(
file_path=file_path)
in_data_features = np.asarray(in_data_features)
voice_activity_detection_data_features = np.asarray(
voice_activity_detection_data_features)
N_frames = in_data_features.shape[0]
hop_size = self._hparams.frame_size // 4
def get_dataset():
dataset = tf.data.Dataset.from_tensor_slices(({
self.FEATURE_1_NAME:
in_data_features,
self.FEATURE_2_NAME:
voice_activity_detection_data_features
}, np.ones_like(in_data_features)))
dataset = dataset.batch(batch_size=1)
print_info(dataset.output_shapes)
return dataset
predict_fn = estimator.predict(input_fn=lambda: get_dataset())
print_info("Shape of in data: {}".format(in_data_features.shape))
print_info("Number of frames for given file: {}".format(N_frames))
embeddings = []
i = 0
for predicted_value in predict_fn:
# print("i = {}".format(i))
"""
TODO:
strange behaviour!
1 wav file = N samples
Eg: N = 600
FramesPerSample=100, BatchSize = 1, NEFF = 129, EMD_K = 30
For each sample the embeddings is of shape [batch_size * frames_per_sample, NEFF, embd_dim].
For prediction batch size is made 1.
Hence the embeddings colapse to [frames_per_sample, NEFF, embd_dim]
1 sample predictions will have `frames_per_sample` outputs
Eg: If input audio file has 75 frames, the prediction will have [7500, NEFF, embd_dim]
"""
embeddings.append(predicted_value)
i += 1
print_info("Number of embeddings predicted for given file: {}".format(
len(embeddings)))
print_error(np.asarray(embeddings).shape)
N_assign = 0
step = 0
for frame_i in tqdm(range(N_frames)):
# expand the dimesion to be inline with TF batch size
in_data_np = np.expand_dims(in_data_features[frame_i], axis=0)
in_phase_np = np.expand_dims(phase_features[frame_i], axis=0)
voice_activity_detection_data_np = np.expand_dims(
voice_activity_detection_data_features[frame_i], axis=0)
embedding_np = np.asarray(
embeddings[frame_i:frame_i + self._hparams.frames_per_sample])
# ----------------------------------------------
embedding_ac = []
for i, j in itertools.product(
range(self._hparams.frames_per_sample),
range(self._hparams.neff)):
if voice_activity_detection_data_np[0, i, j] == 1:
embedding_ac.append(embedding_np[i, j, :])
kmean = KMeans(n_clusters=2, random_state=0).fit(embedding_ac)
# visualization using 3 PCA
pca_Data = PCA(n_components=3).fit_transform(embedding_ac)
fig = plt.figure(1, figsize=(8, 6))
ax = Axes3D(fig, elev=-150, azim=110)
# ax.scatter(pca_Data[:, 0], pca_Data[:, 1], pca_Data[:, 2],
# c=kmean.labels_, cmap=plt.cm.Paired)
ax.scatter(pca_Data[:, 0],
pca_Data[:, 1],
pca_Data[:, 2],
cmap=plt.cm.Paired)
ax.set_title('Embedding visualization using the first 3 PCs')
ax.set_xlabel('1st pc')
ax.set_ylabel('2nd pc')
ax.set_zlabel('3rd pc')
if not os.path.exists("vis"):
os.makedirs("vis")
plt.savefig('vis/' + str(step) + 'pca.jpg')
step += 1
def discriminator(self, images, input_z, reuse=False):
"""
Create the _discriminator network
:param image: Tensor of input image(s)
:param reuse: Boolean if the weights should be reused
:return: Tuple of (tensor output of the _discriminator, tensor logits of the _discriminator)
"""
_, width, height, channel = images.get_shape().as_list()
print_error("{}".format([width, height, channel]))
with tf.variable_scope('_discriminator', reuse=reuse):
# y = tf.reshape(input_z, [-1, 1, 1, 740], name="y_reshape") #2*2*185=>740
# input_z = tf.layers.batch_normalization(input_z)
# input_z = tf.layers.dense(input_z, width*height*channel)
# y = tf.reshape(input_z, [-1, width,height,channel], name="y_reshape")
# print_error(y)
# x1 = conv_cond_concat(y,images)
'''
# c_code = tf.expand_dims(tf.expand_dims(input_z, 1), 1)
# c_code = tf.tile(c_code, [1, 1, 1, 740])
# x1 = tf.concat([images, c_code], 3)
# print_error(x1)
'''
# x1 = tf.layers.batch_normalization(images)
# Input layer consider ?x32x32x3
x1 = tf.layers.conv2d(
images,
64,
5,
strides=2,
padding='same',
kernel_initializer=tf.truncated_normal_initializer(
stddev=0.02))
x1 = tf.maximum(self.alpha * x1, x1)
x1 = tf.layers.batch_normalization(x1, training=True)
# relu1 = tf.layers.dropout(relu1, rate=0.5)
# 16x16x64
# print(x1)
x2 = tf.layers.conv2d(
x1,
128,
5,
strides=2,
padding='same',
kernel_initializer=tf.truncated_normal_initializer(
stddev=0.02))
x2 = tf.maximum(self.alpha * x2, x2)
x2 = tf.layers.batch_normalization(x2, training=True)
# relu2 = tf.layers.dropout(relu2, rate=0.5)
# 8x8x128
# print(x2)
x3 = tf.layers.conv2d(
x2,
256,
5,
strides=2,
padding='same',
kernel_initializer=tf.truncated_normal_initializer(
stddev=0.02))
x3 = tf.maximum(self.alpha * x3, x3)
x3 = tf.layers.batch_normalization(x3, training=True)
# relu3 = tf.layers.dropout(relu3, rate=0.5)
# 4x4x256
# print(x3)
# Flatten it
flat = tf.reshape(x3, (-1, 4 * 4 * 256))
flat = tf.concat([flat, input_z], -1)
# conditioned_fully_connected_layer = tf.concat([flat], axis=-1)
#
# flat = tf.layers.dense(flat, 512)
# flat = tf.layers.dense(flat, 1024)
# flat = tf.layers.dense(flat, 512)
# flat = tf.layers.dense(flat, 256)
# flat = tf.layers.dense(flat, 128)
logits = tf.layers.dense(flat, 512)
logits = tf.maximum(self.alpha * logits, logits)
logits = tf.layers.dense(logits, 1)
# print(logits)
out = tf.sigmoid(logits)
# print('_discriminator out: ', out)
print_info("======>out: {}".format(out))
return out, logits
def predict_on_instance(self, executor, file_path):
"""
Given a mixed audio file, it generates three audio files : mix recontructed, source1, source 2
#TODO debug!!!
:param executor:
:param test_file_path:
:return:
"""
estimator = executor.estimator
in_data_features, voice_activity_detection_data_features, phase_features = self._get_predict_samples(
file_path=file_path)
print_info("in_data_features original shape: {}".format(
in_data_features.shape))
voice_activity_detection_data_features = np.asarray(
voice_activity_detection_data_features)
num_samples_N = in_data_features.shape[0]
hop_size = self._hparams.frame_size // 4 # 256 / 4 = 64
# (38 * 100 -1) * 64 + 256 = 243392 for wav file of 10 seconds appended 3 three times i.e 10 * 8000 (Sampling rate) * = 240000
out_audio1 = np.zeros([
(num_samples_N * self._hparams.frames_per_sample - 1) * hop_size +
self._hparams.frame_size
])
out_audio2 = np.zeros([
(num_samples_N * self._hparams.frames_per_sample - 1) * hop_size +
self._hparams.frame_size
])
mix = np.zeros([
(num_samples_N * self._hparams.frames_per_sample - 1) * hop_size +
self._hparams.frame_size
])
def get_dataset():
dataset = tf.data.Dataset.from_tensor_slices(({
self.FEATURE_1_NAME:
in_data_features,
self.FEATURE_2_NAME:
voice_activity_detection_data_features
}, np.ones_like(in_data_features)))
dataset = dataset.batch(batch_size=1)
print_info(dataset.output_shapes)
return dataset
predict_fn = estimator.predict(input_fn=lambda: get_dataset())
print_info("Shape of in data: {}".format(in_data_features.shape))
print_info("Number of sample for given file: {}".format(num_samples_N))
embeddings = []
i = 0
for predicted_value in predict_fn:
# print("i = {}".format(i))
"""
TODO:
strange behaviour!
1 wav file = N samples
Eg: N = 600
FramesPerSample=100, BatchSize = 1, NEFF = 129, EMD_K = 30
For each sample the embeddings is of shape [batch_size * frames_per_sample, NEFF, embd_dim].
For prediction batch size is made 1.
Hence the embeddings colapse to [frames_per_sample, NEFF, embd_dim]
1 sample predictions will have `frames_per_sample` outputs
Eg: If input audio file has 75 frames, the prediction will have [7500, NEFF, embd_dim]
"""
embeddings.append(predicted_value)
i += 1
# Number of embeddings predicted for given file: 3800 with shape (3800, 129, 40) [number of samples * frames_per_sample, NEFF, EMBD_DIM]
print_info(
"Number of embeddings predicted for given file: {} with shape {}".
format(len(embeddings),
np.asarray(embeddings).shape))
N_assign = 0
# # for every chunk of frames of data
for sample_i in tqdm(range(num_samples_N)): # num_samples = 38
# expand the dimesion to be inline with TF batch size
in_data_np = np.expand_dims(in_data_features[sample_i], axis=0)
in_phase_np = np.expand_dims(phase_features[sample_i], axis=0)
voice_activity_detection_data_np = np.expand_dims(
voice_activity_detection_data_features[sample_i], axis=0)
print_info("in_data_np : ")
print_error(in_data_np.shape)
"""
0*100 to (0+1)*100 = 0 to 100
1*100 to (1+1)*100 = 100 to 200
2*100 to (2+1)*100 = 200 to 300
"""
embedding_np = np.asarray(
embeddings[sample_i *
self._hparams.frames_per_sample:(sample_i + 1) *
self._hparams.frames_per_sample])
# ----------------------------------------------
# embedding_ac = []
# for i, j in itertools.product(range(self._hparams.frames_per_sample), range(self._hparams.neff)):
# if voice_activity_detection_data_np[0, i, j] == 1:
# embedding_ac.append(embedding_np[i, j, :])
embedding_ac = [
embedding_np[i, j, :] for i, j in itertools.product(
range(self._hparams.frames_per_sample),
range(self._hparams.neff))
if voice_activity_detection_data_np[0, i, j] == 1
]
print_error(np.array(embedding_ac).shape)
if embedding_ac == []:
break
kmean = KMeans(n_clusters=2, random_state=0).fit(embedding_ac)
# ----------------------------------------------
mask = np.zeros(
[self._hparams.frames_per_sample, self._hparams.neff, 2])
ind = 0
# print_info("N_assign : {}".format(N_assign))
center = kmean.cluster_centers_
center = center * 0.7 + 0.3 * kmean.cluster_centers_
# two speakers have appeared
# print_info("Found 2 speakers ...")
center_new = kmean.cluster_centers_
cor = np.matmul(center_new[0, :], np.transpose(center))
# rearrange their sequence if not consistant with previous
# frames
# print_info("Correlation : {}".format(cor))
if (cor[1] > cor[0]):
kmean.cluster_centers_ = np.array(
[kmean.cluster_centers_[1], kmean.cluster_centers_[0]])
kmean.labels_ = (kmean.labels_ == 0).astype('int')
kmean.labels_ = (~kmean.labels_).astype('int')
# print_info("center : {}".format(center))
# print_info("kmean.labels_ : {}".format(kmean.labels_))
# ----------------------------------------------
# print_error(kmean.labels_)
# transform the clustering result and voice_activity_detection info. into masks
for i in range(self._hparams.frames_per_sample):
for j in range(self._hparams.neff):
if voice_activity_detection_data_np[0, i, j] == 1:
mask[i, j, kmean.labels_[ind]] = 1
ind += 1
for i in range(self._hparams.frames_per_sample):
# apply the mask and reconstruct the waveform
tot_ind = sample_i * self._hparams.frames_per_sample + i
# ipdb.set_trace()
amp = in_data_np[
0,
i, :] * self._hparams.global_std + self._hparams.global_mean
out_data1 = (mask[i, :, 0] * amp *
voice_activity_detection_data_np[0, i, :])
out_data2 = (mask[i, :, 1] * amp *
voice_activity_detection_data_np[0, i, :])
out_mix = amp
out_data1_l = 10**(out_data1 / 20) / self._hparams.amp_fac
out_data2_l = 10**(out_data2 / 20) / self._hparams.amp_fac
out_mix_l = 10**(out_mix / 20) / self._hparams.amp_fac
out_stft1 = out_data1_l * in_phase_np[0, i, :]
out_stft2 = out_data2_l * in_phase_np[0, i, :]
out_stft_mix = out_mix_l * in_phase_np[0, i, :]
con_data1 = out_stft1[-2:0:-1].conjugate()
con_data2 = out_stft2[-2:0:-1].conjugate()
con_mix = out_stft_mix[-2:0:-1].conjugate()
out1 = np.concatenate((out_stft1, con_data1))
out2 = np.concatenate((out_stft2, con_data2))
out_mix = np.concatenate((out_stft_mix, con_mix))
frame_out1 = np.fft.ifft(out1).astype(np.float64)
frame_out2 = np.fft.ifft(out2).astype(np.float64)
frame_mix = np.fft.ifft(out_mix).astype(np.float64)
start = tot_ind * hop_size
out_audio1[start:(start +
len(frame_out1))] += frame_out1 * 0.5016
out_audio2[start:(start +
len(frame_out2))] += frame_out2 * 0.5016
mix[start:(start + len(frame_mix))] += frame_mix * 0.5016
## the audio has been padded 3 times in AudioReader
## restore the original audio
len1 = len(out_audio1) // 3
len2 = len(out_audio2) // 3
source1 = out_audio1[len1:2 * len1]
source2 = out_audio2[len2:2 * len2]
mix = mix[len2:2 * len2]
# Length of source_1 81130 source_1 shape (81130,)
# Length of source_1 81130 source_1 shape (81130,)
print_info("Length of source_1 {} source_1 shape {}".format(
len1, source1.shape))
print_info("Length of source_1 {} source_1 shape {}".format(
len2, source2.shape))
print_info("Writing file {}".format(
os.path.splitext(file_path)[0] + "_source1.wav"))
print_info("Writing file {}".format(
os.path.splitext(file_path)[0] + "_source2.wav"))
librosa.output.write_wav(
os.path.splitext(file_path)[0] + "_source1.wav", source1,
self._hparams.sampling_rate)
librosa.output.write_wav(
os.path.splitext(file_path)[0] + "_source2.wav", source2,
self._hparams.sampling_rate)
librosa.output.write_wav(
os.path.splitext(file_path)[0] + "_full.wav", mix,
self._hparams.sampling_rate)
return [(source1, self._hparams.sampling_rate),
(source2, self._hparams.sampling_rate)]