def create_feature_files(data_path, output_path, df_files, feat_config):
files = df_files.copy()
# Extract the feature settings
wlen = feat_config['window_length']
wstep = feat_config['window_step']
srate = feat_config['sample_rate']
fsize = feat_config['feature_size']
# load all the files needer per subjects
l = df_files.shape[0]
for i in range(0, l):
print('Processing file ', i + 1, ' of ', l)
file_name = df_files.at[i, 'FileName']
feat_file = savepath = os.path.splitext(file_name)[0] + '.npy'
# DELTA function that extracts from the file the features
# the resulting data gets dumped to an equivalently named
# feature numpy file.
speech_feature.extract_feature(file_name,
winlen=wlen,
winstep=wstep,
sr=srate,
feature_size=fsize,
feature_name='fbank')
# Move the file to the feature path
dest_file = output_path + '/' + os.path.basename(feat_file)
data = np.load(feat_file)
os.rename(feat_file, dest_file)
# Replace the file name with the feature file name
files.at[i, 'FileName'] = dest_file
return files
def test_tf_delta_detla(self):
''' test tensorflow delta delta '''
speech_feature.extract_feature((self.wavfile),
winlen=self.winlen,
winstep=self.winstep,
sr=self.sr,
feature_size=self.feature_size,
feature_name='fbank')
feat = np.load(self.featfile)
self.assertEqual(feat.shape, (425, 40, 1))
feat = speech_feature.add_delta_delta(feat, 40, order=2)
self.assertEqual(feat.shape, (425, 40, 3))
def test_tf_spec(self):
''' test tensorflow spec feature interface '''
speech_feature.extract_feature((self.wavfile),
winlen=self.winlen,
winstep=self.winstep,
sr=self.sr,
feature_size=self.feature_size,
feature_name='spec')
feat = np.load(self.featfile)
self.assertEqual(feat.shape, (425, 129, 1))
with self.session(use_gpu=False):
feat = speech_ops.delta_delta(feat, 2)
self.assertEqual(feat.eval().shape, (425, 129, 3))
def test_tf_fbank(self):
''' test tensorflow fbank feature interface '''
speech_feature.extract_feature((self.wavfile),
winlen=self.winlen,
winstep=self.winstep,
sr=self.sr,
feature_size=self.feature_size,
feature_name='fbank')
feat = np.load(self.featfile)
logging.info(f"feat : {feat}")
self.assertEqual(feat.shape, (425, 40, 1))
with self.session(use_gpu=False):
feat = speech_ops.delta_delta(feat, 2)
self.assertEqual(feat.eval().shape, (425, 40, 3))