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_delta_delta(self):
''' test add delta detlas '''
#pylint: disable=invalid-name
p = tffeat.speech_params(sr=self.sr_true,
bins=40,
cmvn=False,
audio_desired_samples=1000,
add_delta_deltas=False)
with self.session():
wavfile = tf.constant(self.wavpath)
audio, sample_rate = tffeat.read_wav(wavfile, self.hp)
del sample_rate
feature = tffeat.compute_mel_filterbank_features(
audio,
sample_rate=p.audio_sample_rate,
preemphasis=p.audio_preemphasis,
frame_length=p.audio_frame_length,
frame_step=p.audio_frame_step,
lower_edge_hertz=p.audio_lower_edge_hertz,
upper_edge_hertz=p.audio_upper_edge_hertz,
num_mel_bins=p.audio_num_mel_bins,
apply_mask=False)
feature = tffeat.delta_delta(feature, order=2)
self.assertEqual(feature.eval().shape, (11, 40, 3))
def load_wav(wavpath, sr=8000):
'''
audio:
np.float32, shape [None], sample in [-1, 1], using librosa.load
np.int16, shape [None], sample in [-32768, 32767], using scipy.io.wavfile
np.float32, shape[None, audio_channel], sample int [-1, 1], using tf.DecodeWav
return
sr: sample rate
audio: [-1, 1], same to tf.DecodeWav
'''
#from scipy.io import wavfile
#sample_rate, audio = wavfile.read(wavpath)
#samples, sample_rate = librosa.load(wavpath, sr=sr)
with tf.Session() as sess:
params = speech_ops.speech_params(sr=sr, audio_desired_samples=-1)
t_wavpath = tf.placeholder(dtype=tf.string)
t_audio, t_sample_rate = speech_ops.read_wav(t_wavpath, params)
audio, sample_rate = sess.run([t_audio, t_sample_rate],
feed_dict={t_wavpath: wavpath})
audio = audio[:, 0]
assert sample_rate == sr, 'sampling rate must be {}Hz, get {}Hz'.format(
sr, sample_rate)
return sample_rate, audio
def test_read_wav(self):
''' test read wav op '''
with self.session():
wavfile = tf.constant(self.wavpath)
# read wav
audio, sample_rate = tffeat.read_wav(wavfile, self.hp)
self.assertEqual(sample_rate.eval(), self.sr_true)
self.assertEqual(audio.eval().shape, (1000, 1))
self.assertAllEqual(audio.eval()[:, 0], self.audio_true[:1000])
def test_extract_logfbank_with_delta(self):
''' test logfbank with delta op'''
#pylint: disable=invalid-name
hp = tffeat.speech_params(sr=self.sr_true,
bins=40,
cmvn=False,
audio_desired_samples=1000,
add_delta_deltas=False)
with self.session():
wavfile = tf.constant(self.wavpath)
# read wav
audio, sample_rate = tffeat.read_wav(wavfile, hp)
del sample_rate
# fbank with delta delta
fbank = tffeat.extract_logfbank_with_delta(audio, hp)
self.assertEqual(fbank.eval().shape, (11, 40, 1))
def extract_filterbank(*args, **kwargs):
''' tensorflow fbank feat '''
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')
dry_run = kwargs.get('dry_run')
del nfft
feat_name = 'fbank'
graph = None
op = None
# get session
if feat_name not in _global_sess:
graph = tf.Graph()
#pylint: disable=not-context-manager
with graph.as_default():
# fbank
params = speech_ops.speech_params(
sr=sr,
bins=feature_size,
add_delta_deltas=False,
audio_frame_length=winlen,
audio_frame_step=winstep)
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)
fbank = tf.identity(fbank, name=feat_name)
sess = _get_session(_get_out_tensor_name(feat_name, 0), graph)
for wavpath in args:
savepath = os.path.splitext(wavpath)[0] + '.npy'
logging.debug('input: {}, output: {}'.format(wavpath, savepath))
feat = sess.run(feat_name + ":0", feed_dict={'wavpath:0': wavpath})
# save feat
if dry_run:
logging.info('save feat: path {} shape:{} dtype:{}'.format(
savepath, feat.shape, feat.dtype))
else:
np.save(savepath, feat)
def test_extract_feature(self):
''' test logfbank with delta, and cmvn '''
#pylint: disable=invalid-name
hp = tffeat.speech_params(sr=self.sr_true,
bins=40,
cmvn=False,
audio_desired_samples=1000,
add_delta_deltas=True)
with self.cached_session(use_gpu=False, force_gpu=False):
wavfile = tf.constant(self.wavpath)
# read wav
audio, sample_rate = tffeat.read_wav(wavfile, hp)
del sample_rate
# fbank with delta delta and cmvn
feature = tffeat.extract_feature(audio, hp)
self.assertEqual(feature.eval().shape, (11, 40, 3))
def test_powspec_feat(self):
''' test spectrogram op '''
with self.cached_session(use_gpu=False, force_gpu=False):
wavfile = tf.constant(self.wavpath)
# read wav
audio, sample_rate = tffeat.read_wav(wavfile, self.hp)
del sample_rate
# spectorgram
spectrogram = tffeat.powspec_feat(
audio,
sr=self.sr_true,
nfft=None,
winlen=self.hp.audio_frame_length,
winstep=self.hp.audio_frame_step,
lowfreq=self.hp.audio_lower_edge_hertz,
highfreq=self.hp.audio_upper_edge_hertz,
preemph=self.hp.audio_preemphasis)
nfft = int(np.log2(self.hp.audio_frame_length * self.sr_true)) + 1
nfft = 1 << nfft
self.assertEqual(spectrogram.eval().shape, (1, 11, int(nfft / 2 + 1)))
def test_batch_extract_feature(self):
''' test batched feature extraction '''
#pylint: disable=invalid-name
hp = tffeat.speech_params(sr=self.sr_true,
bins=40,
cmvn=False,
audio_desired_samples=1000,
add_delta_deltas=True)
batch_size = 2
with self.session():
wavfile = tf.constant(self.wavpath)
# read wav
audio, sample_rate = tffeat.read_wav(wavfile, hp)
del sample_rate
audio = tf.stack([audio] * batch_size)
# fbank with delta delta and cmvn
feature = tffeat.batch_extract_feature(audio, hp)
self.assertEqual(feature.eval().shape, (batch_size, 11, 40, 3))
def load_wav(wavpath, sr=8000):
'''
audio:
np.float32, shape [None], sample in [-1, 1], using librosa.load
np.int16, shape [None], sample in [-32768, 32767], using scipy.io.wavfile
np.float32, shape[None, audio_channel], sample int [-1, 1], using tf.DecodeWav
return
sr: sample rate
audio: [-1, 1], same to tf.DecodeWav
'''
#from scipy.io import wavfile
#sample_rate, audio = wavfile.read(wavpath)
#samples, sample_rate = librosa.load(wavpath, sr=sr)
feat_name = 'load_wav'
graph = None
# get session
if feat_name not in _global_sess:
graph = tf.Graph()
with graph.as_default():
params = speech_ops.speech_params(sr=sr, audio_desired_samples=-1)
t_wavpath = tf.placeholder(dtype=tf.string, name="wavpath")
t_audio, t_sample_rate = speech_ops.read_wav(t_wavpath, params)
t_audio = tf.identity(t_audio, name="audio")
t_sample_rate = tf.identity(t_sample_rate, name="sample_rate")
sess = _get_session(feat_name, graph)
audio, sample_rate = sess.run([
_get_out_tensor_name('audio', 0),
_get_out_tensor_name('sample_rate', 0)
],
feed_dict={"wavpath:0": wavpath})
audio = audio[:, 0]
assert sample_rate == sr, 'sampling rate must be {}Hz, get {}Hz'.format(
sr, sample_rate)
return sample_rate, audio
