def extract(x, sr=16000):
f_global = []
# 34D short-term feature
f = ShortTermFeatures.feature_extraction(x, sr, globalvars.frame_size * sr,
globalvars.step * sr)
# for pyAudioAnalysis which support python3
if type(f) is tuple:
f = f[0]
# Harmonic ratio and pitch, 2D
hr_pitch = ShortTermFeatures.speed_feature(x, int(sr),
int(globalvars.frame_size * sr),
int(globalvars.step * sr))
f = np.append(f, hr_pitch.transpose(), axis=0)
# Z-normalized
f = stats.zscore(f, axis=0)
f = f.transpose()
f_global.append(f)
f_global = sequence.pad_sequences(
f_global,
maxlen=globalvars.max_len,
dtype="float32",
padding="post",
value=globalvars.masking_value,
)
return f_global
def _get_batches_of_transformed_samples(self, index_array):
batch_x = []
for i, j in enumerate(index_array):
x = self.x[j]
# Augmentation
if self.audio_data_generator.white_noise_:
x = self.audio_data_generator.white_noise(x)
if self.audio_data_generator.shift_:
x = self.audio_data_generator.shift(x)
if self.audio_data_generator.stretch_:
x = self.audio_data_generator.stretch(x)
# 34D short-term feature
f = ShortTermFeatures.feature_extraction(
x, self.sr, globalvars.frame_size * self.sr,
globalvars.step * self.sr)
# Harmonic ratio and pitch, 2D
hr_pitch = ShortTermFeatures.speed_feature(
x, self.sr, globalvars.frame_size * self.sr,
globalvars.step * self.sr)
x = np.append(f, hr_pitch.transpose(), axis=0)
# Z-normalized
x = stats.zscore(x, axis=0)
x = x.transpose()
batch_x.append(x)
batch_x = sequence.pad_sequences(
batch_x,
maxlen=globalvars.max_len,
dtype="float32",
padding="post",
value=globalvars.masking_value,
)
batch_u = np.full(
(
len(index_array),
globalvars.nb_attention_param,
),
globalvars.attention_init_value,
dtype=np.float32,
)
if self.y is None:
return [batch_u, batch_x]
batch_y = self.y[index_array]
return [batch_u, batch_x], batch_y
def extract_dataset(data, nb_samples, dataset, save=True):
f_global = []
i = 0
for (x, Fs) in data:
# 34D short-term feature
f = ShortTermFeatures.feature_extraction(x, Fs,
globalvars.frame_size * Fs,
globalvars.step * Fs)
# for pyAudioAnalysis which support python3
if type(f) is tuple:
f = f[0]
# Harmonic ratio and pitch, 2D
hr_pitch = ShortTermFeatures.speed_feature(x, Fs,
globalvars.frame_size * Fs,
globalvars.step * Fs)
f = np.append(f, hr_pitch.transpose(), axis=0)
# Z-normalized
f = stats.zscore(f, axis=0)
f = f.transpose()
f_global.append(f)
sys.stdout.write("\033[F")
i = i + 1
print("Extracting features " + str(i) + "/" + str(nb_samples) +
" from data set...")
f_global = sequence.pad_sequences(
f_global,
maxlen=globalvars.max_len,
dtype="float32",
padding="post",
value=globalvars.masking_value,
)
if save:
print("Saving features to file...")
pickle.dump(f_global, open(dataset + "_features.p", "wb"))
return f_global
