def call(self, audio_data, sample_rate=None):
"""
Caculate power spectrum or log power spectrum of audio data.
:param audio_data: the audio signal from which to compute spectrum. Should be an (1, N) tensor.
:param sample_rate: [option]the samplerate of the signal we working with, default is 16kHz.
:return: A float tensor of size (num_frames, num_frequencies) containing power spectrum (output_type=1)
or log power spectrum (output_type=2) of every frame in speech.
"""
p = self.config
with tf.name_scope('spectrum'):
if sample_rate == None:
sample_rate = tf.constant(p.sample_rate, dtype=tf.int32)
assert_op = tf.assert_equal(tf.constant(p.sample_rate),
tf.cast(sample_rate, dtype=tf.int32))
with tf.control_dependencies([assert_op]):
sample_rate = tf.cast(sample_rate, dtype=float)
spectrum = py_x_ops.spectrum(
audio_data,
sample_rate,
window_length=p.window_length,
frame_length=p.frame_length,
output_type=p.output_type,
snip_edges=p.snip_edges,
raw_energy=p.raw_energy,
preEph_coeff=p.preeph_coeff,
window_type=p.window_type,
remove_dc_offset=p.remove_dc_offset,
is_fbank=p.is_fbank)
return spectrum
def _freq_feat_graph(feat_name, **kwargs):
winlen = kwargs.get('winlen')
winstep = kwargs.get('winstep')
feature_size = kwargs.get('feature_size')
sr = kwargs.get('sr') #pylint: disable=invalid-name
nfft = kwargs.get('nfft')
del nfft
assert feat_name in ('fbank', 'spec')
params = speech_ops.speech_params(sr=sr,
bins=feature_size,
add_delta_deltas=False,
audio_frame_length=winlen,
audio_frame_step=winstep)
graph = None
if feat_name == 'fbank':
# get session
if feat_name not in _global_sess:
graph = tf.Graph()
#pylint: disable=not-context-manager
with graph.as_default():
# fbank
filepath = tf.placeholder(dtype=tf.string,
shape=[],
name='wavpath')
waveforms, sample_rate = speech_ops.read_wav(filepath, params)
del sample_rate
fbank = speech_ops.extract_feature(waveforms, params)
# shape must be [T, D, C]
feat = tf.identity(fbank, name=feat_name)
elif feat_name == 'spec':
# magnitude spec
if feat_name not in _global_sess:
graph = tf.Graph()
#pylint: disable=not-context-manager
with graph.as_default():
filepath = tf.placeholder(dtype=tf.string,
shape=[],
name='wavpath')
waveforms, sample_rate = speech_ops.read_wav(filepath, params)
spec = py_x_ops.spectrum(
waveforms[:, 0],
tf.cast(sample_rate, tf.dtypes.float32),
output_type=1
) #output_type: 1, power spec; 2 log power spec
spec = tf.sqrt(spec)
# shape must be [T, D, C]
spec = tf.expand_dims(spec, -1)
feat = tf.identity(spec, name=feat_name)
else:
raise ValueError(f"Not support freq feat: {feat_name}.")
return graph, (_get_out_tensor_name('wavpath',
0), _get_out_tensor_name(feat_name, 0))
def test_spectrum(self):
''' test spectrum op'''
with self.session():
sample_rate, input_data = feat_lib.load_wav(self.wavpath, sr=16000)
output = py_x_ops.spectrum(input_data, sample_rate)
output_true = np.array(
[[-16.863441, -16.910473, -17.077059, -16.371634, -16.845686],
[-17.922068, -20.396345, -19.396944, -17.331493, -16.118851],
[-17.017776, -17.551350, -20.332376, -17.403994, -16.617926],
[-19.873854, -17.644503, -20.679525, -17.093716, -16.535091],
[-17.074402, -17.295971, -16.896650, -15.995432, -16.560730]])
self.assertEqual(tf.rank(output).eval(), 2)
logging.info('Shape of spectrum: {}'.format(output.shape))
self.assertAllClose(output.eval()[4:9, 4:9], output_true)
def test_spectrum(self):
''' test spectrum op'''
with self.session():
sample_rate, input_data = feat_lib.load_wav(self.wavpath, sr=16000)
output = py_x_ops.spectrum(input_data, sample_rate)
output_true = np.array([
-16.018925, -16.491777, -16.903442, -18.108875, -19.477205,
-19.039738, -17.066263, -16.530647, -16.033670, -15.492795,
-15.347169, -16.443783, -15.385968, -15.631793, -16.286760,
-16.555447, -15.107640, -15.158586, -16.397518, -14.803325,
-15.173873, -15.785010, -15.551179, -15.487743, -15.732930,
-15.610220, -15.314099, -14.765355, -14.572725, -13.482535,
-13.463938, -14.457010, -16.253452, -15.444997, -13.472414,
-12.852523, -13.163157, -13.957175, -14.148843, -13.527264,
-12.840333, -13.056757, -14.582790, -13.900843, -13.864534,
-14.037180, -15.386706, -16.500109, -16.309618, -13.585808
])
self.assertEqual(tf.rank(output).eval(), 1)
self.assertAllClose(output.eval().flatten()[:50], output_true)
def test_spectrum(self):
''' test spectrum op'''
with self.cached_session(use_gpu=False, force_gpu=False):
sample_rate, input_data = feat_lib.load_wav(self.wavpath, sr=16000)
logging.info(
f"input shape: {input_data.shape}, sample rate dtype: {sample_rate.dtype}"
)
self.assertEqual(sample_rate, 16000)
output = py_x_ops.spectrum(input_data, sample_rate)
#pylint: disable=bad-whitespace
output_true = np.array(
[[-16.863441, -16.910473, -17.077059, -16.371634, -16.845686],
[-17.922068, -20.396345, -19.396944, -17.331493, -16.118851],
[-17.017776, -17.551350, -20.332376, -17.403994, -16.617926],
[-19.873854, -17.644503, -20.679525, -17.093716, -16.535091],
[-17.074402, -17.295971, -16.896650, -15.995432, -16.560730]])
#pylint: enable=bad-whitespace
self.assertEqual(tf.rank(output).eval(), 2)
logging.info('Shape of spectrum: {}'.format(output.shape))
self.assertAllClose(output.eval()[4:9, 4:9], output_true)