def test_load_audio(self):
indexes = np.arange(5)
x_data_raw, y_data_raw, sr = load_audio(self.df,
indexes_in_batch=indexes)
self.assertEqual(len(x_data_raw), 5)
self.assertEqual(len(y_data_raw), 5)
def test_extract_features_and_pad(self):
indexes = np.arange(5)
x_data_raw, y_data_raw, sr = load_audio(self.df,
indexes_in_batch=indexes)
x_data, input_length = self.dg.extract_features_and_pad(x_data_raw, sr)
self.assertEqual(x_data.shape, (5, 382, 26))
self.assertEqual(len(input_length), 5)
self.assertLessEqual(all(input_length), 382)
def test_convert_transcripts(self):
_, y_data_raw, sr = load_audio(self.df, indexes_in_batch=[0])
transcript, y_length = convert_and_pad_transcripts(y_data_raw)
exp = [
23., 5., 18., 5., 0., 9., 0., 2., 21., 20., 0., 1., 12., 18., 5.,
1., 4., 25., 0., 15., 14., 0., 20., 8., 5., 0., 3., 1., 18., 20.
]
list = transcript[0].tolist()
self.assertListEqual(list, exp)
self.assertEqual(y_length, 30)
def test_extract_mel_spec(self):
x_data_raw, _, sr = load_audio(self.df, indexes_in_batch=[0])
mel_spec, x_length = extract_mel_spectrogram_and_pad(
x_data_raw[0],
sr=sr,
max_pad_length=500,
frame_length=320,
hop_length=160,
n_mels=40)
self.assertTupleEqual(mel_spec.shape, (500, 40))
self.assertEqual(x_length, 256)
def test_extract_mfcc(self):
x_data_raw, _, sr = load_audio(self.df, indexes_in_batch=[0])
mfcc_padded, x_length = extract_mfcc_and_pad(x_data_raw[0],
sr=sr,
max_pad_length=500,
frame_length=320,
hop_length=160,
mfcc_features=26,
n_mels=40)
self.assertTupleEqual(mfcc_padded.shape, (500, 26))
self.assertEqual(x_length, 256)
def __getitem__(self, batch_index):
"""
Generates a batch of correctly shaped X and Y data
:param batch_index: index of the batch to generate
:return: input dictionary containing:
'the_input': np.ndarray[shape=(batch_size, max_seq_length, mfcc_features)]: input audio data
'the_labels': np.ndarray[shape=(batch_size, max_transcript_length)]: transcription data
'input_length': np.ndarray[shape=(batch_size, 1)]: length of each sequence (numb of frames) in x_data
'label_length': np.ndarray[shape=(batch_size, 1)]: length of each sequence (numb of letters) in y_data
output dictionary containing:
'ctc': np.ndarray[shape=(batch_size, 1)]: dummy data for dummy loss function
"""
# Generate indexes of current batch
indexes_in_batch = self.indexes[batch_index *
self.batch_size:(batch_index + 1) *
self.batch_size]
# Shuffle indexes within current batch if shuffle=true
if self.shuffle:
shuf(indexes_in_batch)
# Load audio and transcripts
x_data_raw, y_data_raw, sr = load_audio(self.df, indexes_in_batch)
# Preprocess and pad data
x_data, input_length = self.extract_features_and_pad(x_data_raw, sr)
y_data, label_length = convert_and_pad_transcripts(y_data_raw)
# print ("\nx_data shape: ", x_data.shape)
# print ("y_data shape: ", y_data.shape)
# print ("input_length shape: ", input_length.shape)
# print ("label_length shape: ", label_length.shape)
# print ("input length: ", input_length)
# print ("label_length: ", label_length, "\n")
inputs = {
'the_input': x_data,
'the_labels': y_data,
'input_length': input_length,
'label_length': label_length
}
outputs = {
'ctc': np.zeros([self.batch_size])
} # dummy data for dummy loss function
return inputs, outputs
def test_get_seq_size(self):
x_data_raw, _, sr = load_audio(self.df, indexes_in_batch=[0])
size = self.dg.get_seq_size(x_data_raw[0], sr)
self.assertEqual(size, 256)