def samples_to_mfccs_orig(samples, sample_rate, train_phase=False):
#tf.print('window_size: ', Config.audio_window_samples, ' stride: ', Config.audio_step_samples)
spectrogram = contrib_audio.audio_spectrogram(samples,
window_size=Config.audio_window_samples,
stride=Config.audio_step_samples,
magnitude_squared=True)
# Data Augmentations
if train_phase:
if FLAGS.augmentation_spec_dropout_keeprate < 1:
spectrogram = augment_dropout(spectrogram,
keep_prob=FLAGS.augmentation_spec_dropout_keeprate)
if FLAGS.augmentation_freq_and_time_masking:
spectrogram = augment_freq_time_mask(spectrogram,
frequency_masking_para=FLAGS.augmentation_freq_and_time_masking_freq_mask_range,
time_masking_para=FLAGS.augmentation_freq_and_time_masking_time_mask_range,
frequency_mask_num=FLAGS.augmentation_freq_and_time_masking_number_freq_masks,
time_mask_num=FLAGS.augmentation_freq_and_time_masking_number_time_masks)
if FLAGS.augmentation_pitch_and_tempo_scaling:
spectrogram = augment_pitch_and_tempo(spectrogram,
max_tempo=FLAGS.augmentation_pitch_and_tempo_scaling_max_tempo,
max_pitch=FLAGS.augmentation_pitch_and_tempo_scaling_max_pitch,
min_pitch=FLAGS.augmentation_pitch_and_tempo_scaling_min_pitch)
if FLAGS.augmentation_speed_up_std > 0:
spectrogram = augment_speed_up(spectrogram, speed_std=FLAGS.augmentation_speed_up_std)
mfccs = contrib_audio.mfcc(spectrogram, sample_rate, dct_coefficient_count=Config.n_input)
mfccs = tf.reshape(mfccs, [-1, Config.n_input])
#tf.print('dct_count: ', Config.n_input)
return mfccs, tf.shape(input=mfccs)[0]
def samples_to_mfccs(samples, sample_rate, train_phase=False, sample_id=None):
if train_phase:
# We need the lambdas to make TensorFlow happy.
# pylint: disable=unnecessary-lambda
tf.cond(tf.math.not_equal(sample_rate, FLAGS.audio_sample_rate),
lambda: tf.print('WARNING: sample rate of sample', sample_id, '(', sample_rate, ') '
'does not match FLAGS.audio_sample_rate. This can lead to incorrect results.'),
lambda: tf.no_op(),
name='matching_sample_rate')
spectrogram = contrib_audio.audio_spectrogram(samples,
window_size=Config.audio_window_samples,
stride=Config.audio_step_samples,
magnitude_squared=True)
# Data Augmentations
if train_phase:
if FLAGS.augmentation_spec_dropout_keeprate < 1:
spectrogram = augment_dropout(spectrogram,
keep_prob=FLAGS.augmentation_spec_dropout_keeprate)
# sparse warp must before freq/time masking
if FLAGS.augmentation_sparse_warp:
spectrogram = augment_sparse_warp(spectrogram,
time_warping_para=FLAGS.augmentation_sparse_warp_time_warping_para,
interpolation_order=FLAGS.augmentation_sparse_warp_interpolation_order,
regularization_weight=FLAGS.augmentation_sparse_warp_regularization_weight,
num_boundary_points=FLAGS.augmentation_sparse_warp_num_boundary_points,
num_control_points=FLAGS.augmentation_sparse_warp_num_control_points)
if FLAGS.augmentation_freq_and_time_masking:
spectrogram = augment_freq_time_mask(spectrogram,
frequency_masking_para=FLAGS.augmentation_freq_and_time_masking_freq_mask_range,
time_masking_para=FLAGS.augmentation_freq_and_time_masking_time_mask_range,
frequency_mask_num=FLAGS.augmentation_freq_and_time_masking_number_freq_masks,
time_mask_num=FLAGS.augmentation_freq_and_time_masking_number_time_masks)
if FLAGS.augmentation_pitch_and_tempo_scaling:
spectrogram = augment_pitch_and_tempo(spectrogram,
max_tempo=FLAGS.augmentation_pitch_and_tempo_scaling_max_tempo,
max_pitch=FLAGS.augmentation_pitch_and_tempo_scaling_max_pitch,
min_pitch=FLAGS.augmentation_pitch_and_tempo_scaling_min_pitch)
if FLAGS.augmentation_speed_up_std > 0:
spectrogram = augment_speed_up(spectrogram, speed_std=FLAGS.augmentation_speed_up_std)
mfccs = contrib_audio.mfcc(spectrogram=spectrogram,
sample_rate=sample_rate,
dct_coefficient_count=Config.n_input,
upper_frequency_limit=FLAGS.audio_sample_rate/2)
mfccs = tf.reshape(mfccs, [-1, Config.n_input])
return mfccs, tf.shape(input=mfccs)[0]
def samples_to_mfccs(samples, sample_rate, train_phase=False):
spectrogram = contrib_audio.audio_spectrogram(samples,
window_size=Config.audio_window_samples,
stride=Config.audio_step_samples,
magnitude_squared=True)
# Data Augmentations
if train_phase:
if FLAGS.augmentation_spec_dropout_keeprate < 1:
spectrogram = augment_dropout(spectrogram,
keep_prob=FLAGS.augmentation_spec_dropout_keeprate)
# sparse warp must before freq/time masking
if FLAGS.augmentation_sparse_warp:
spectrogram = augment_sparse_warp(spectrogram,
time_warping_para=FLAGS.augmentation_sparse_warp_time_warping_para,
interpolation_order=FLAGS.augmentation_sparse_warp_interpolation_order,
regularization_weight=FLAGS.augmentation_sparse_warp_regularization_weight,
num_boundary_points=FLAGS.augmentation_sparse_warp_num_boundary_points,
num_control_points=FLAGS.augmentation_sparse_warp_num_control_points)
if FLAGS.augmentation_freq_and_time_masking:
spectrogram = augment_freq_time_mask(spectrogram,
frequency_masking_para=FLAGS.augmentation_freq_and_time_masking_freq_mask_range,
time_masking_para=FLAGS.augmentation_freq_and_time_masking_time_mask_range,
frequency_mask_num=FLAGS.augmentation_freq_and_time_masking_number_freq_masks,
time_mask_num=FLAGS.augmentation_freq_and_time_masking_number_time_masks)
if FLAGS.augmentation_pitch_and_tempo_scaling:
spectrogram = augment_pitch_and_tempo(spectrogram,
max_tempo=FLAGS.augmentation_pitch_and_tempo_scaling_max_tempo,
max_pitch=FLAGS.augmentation_pitch_and_tempo_scaling_max_pitch,
min_pitch=FLAGS.augmentation_pitch_and_tempo_scaling_min_pitch)
if FLAGS.augmentation_speed_up_std > 0:
spectrogram = augment_speed_up(spectrogram, speed_std=FLAGS.augmentation_speed_up_std)
mfccs = contrib_audio.mfcc(spectrogram=spectrogram,
sample_rate=sample_rate,
dct_coefficient_count=Config.n_input,
upper_frequency_limit=FLAGS.audio_sample_rate/2)
mfccs = tf.reshape(mfccs, [-1, Config.n_input])
return mfccs, tf.shape(input=mfccs)[0]