def tflite_compat_mel_from_samples(samples, hparams):
"""EXPERIMENTAL: Log mel spec with ops that can be made TFLite compatible."""
features = melspec_input.build_mel_calculation_graph(
samples,
hparams.sample_rate,
window_length_seconds=2048 / hparams.sample_rate,
hop_length_seconds=(hparams.spec_hop_length / hparams.sample_rate),
num_mel_bins=hparams.spec_n_bins,
lower_edge_hz=hparams.spec_fmin,
upper_edge_hz=hparams.sample_rate / 2.0,
frame_width=1,
frame_hop=1,
tflite_compatible=False) # False here, but would be True on device.
return tf.squeeze(features, 1)
def BuildTfGraph(self, tflite_compatible=False):
"""Setup the TF graph using the single function under test."""
if tflite_compatible:
# tflite requires explicit input sizing.
input_length = len(self._test_waveform)
else:
input_length = None
with self._graph.as_default():
waveform_input = tf.placeholder(tf.float32, [input_length])
# This is the single function provided by the library.
features = melspec_input.build_mel_calculation_graph(
waveform_input, tflite_compatible=tflite_compatible)
self._input = waveform_input
self._output = features
def tflite_compat_mel(wav_audio, hparams):
"""EXPERIMENTAL: Log mel spec with ops that can be made TFLite compatible."""
samples, decoded_sample_rate = tf.audio.decode_wav(
wav_audio, desired_channels=1)
samples = tf.squeeze(samples, axis=1)
with tf.control_dependencies(
[tf.assert_equal(decoded_sample_rate, MELSPEC_SAMPLE_RATE)]):
features = melspec_input.build_mel_calculation_graph(
samples, MELSPEC_SAMPLE_RATE,
window_length_seconds=2048 / MELSPEC_SAMPLE_RATE, # 0.128
hop_length_seconds=(
hparams.spec_hop_length / MELSPEC_SAMPLE_RATE), # 0.032
num_mel_bins=hparams.spec_n_bins,
lower_edge_hz=hparams.spec_fmin,
upper_edge_hz=MELSPEC_SAMPLE_RATE / 2.0,
frame_width=1,
frame_hop=1,
tflite_compatible=False) # False here, but would be True on device.
return tf.squeeze(features, 1)
def tflite_compat_mel_from_samples(samples, hparams):
"""EXPERIMENTAL: Log mel spec with ops that can be made TFLite compatible."""
# Ensure hparams.sample_rate is MELSPEC_SAMPLE_RATE because that is what these
# parameters are hard coded to expect.
with tf.control_dependencies(
[tf.assert_equal(hparams.sample_rate, MELSPEC_SAMPLE_RATE)]):
features = melspec_input.build_mel_calculation_graph(
samples,
MELSPEC_SAMPLE_RATE,
window_length_seconds=2048 / MELSPEC_SAMPLE_RATE, # 0.128
hop_length_seconds=(hparams.spec_hop_length /
MELSPEC_SAMPLE_RATE), # 0.032
num_mel_bins=hparams.spec_n_bins,
lower_edge_hz=hparams.spec_fmin,
upper_edge_hz=MELSPEC_SAMPLE_RATE / 2.0,
frame_width=1,
frame_hop=1,
tflite_compatible=False
) # False here, but would be True on device.
return tf.squeeze(features, 1)