def compute_zcr(self):
self.zcr = []
for i in range(0, self.frames):
current_frame = utils._get_frame(self.audio, i, self.frame)
self.zcr.append(
np.mean(0.5 * np.abs(np.diff(np.sign(current_frame)))))
self.zcr = np.asarray(self.zcr)
def compute_mfcc(self):
self.melspectrogram = []
self.logamplitude = []
self.mfcc = []
self.mfcc_delta = []
self.mfcc_delta2 = []
for i in range(0, self.frames-1):
current_frame = utils._get_frame(self.audio, i, self.frame)
# MFCC computation with default settings (2048 FFT window length)
self.melspectrogram.append(librosa.feature.melspectrogram(current_frame, sr=self.sampling_rate,
hop_length=self.frame)[0:,][0:,1])
self.logamplitude.append(librosa.logamplitude(self.melspectrogram[i]))
self.mfcc.append(librosa.feature.mfcc(S=self.logamplitude[i], n_mfcc=self.number_of_mfcc).transpose())
# plt.figure(figsize=(10, 4))
# librosa.display.specshow(self.mfcc[i], x_axis='time')
# plt.colorbar()
# plt.title('MFCC')
# plt.tight_layout()
self.mfcc_delta.append(librosa.feature.delta(self.mfcc[i]))
self.mfcc_delta2.append(librosa.feature.delta(self.mfcc[i], order=2))
self.logamplitude[i]=(self.logamplitude[i].T.flatten()[:, np.newaxis].T)
self.melspectrogram = np.asarray(self.melspectrogram)
self.logamplitude = np.asarray(self.logamplitude)
self.mfcc = np.asarray(self.mfcc)
self.mfcc_delta = np.asarray(self.mfcc_delta)
self.mfcc_delta2 = np.asarray(self.mfcc_delta2)
def compute_energy(self, frame=2048, sampleing_rate=44000):
self.energy = []
for i in range(0, self.frames):
current_frame = utils._get_frame(self.audio, i, frame)
self.energy.append(
np.sum(current_frame**2) / np.float64(len(current_frame)))
self.energy = np.asarray(self.energy)
def compute_geometric_mean(self):
self.geometric_mean = []
for i in range(0, self.frames):
current_frame = utils._get_frame(self.audio, i, self.frame)
sum = np.sum(current_frame**2) # total frame energy
frame_length = len(current_frame)
self.geometric_mean.append(sum / frame_length)
self.geometric_mean = np.asarray(self.geometric_mean)
def compute_energy_entropy(self):
numOfShortBlocks = 10
eps = 0.00000001
self.energy_entropy = []
for i in range(0, self.frames):
current_frame = utils._get_frame(self.audio, i, self.frame)
Eol = np.sum(current_frame**2) # total frame energy
L = len(current_frame)
subWinLength = int(np.floor(L / numOfShortBlocks))
if L != subWinLength * numOfShortBlocks:
current_frame = current_frame[0:subWinLength *
numOfShortBlocks]
# subWindows is of size [numOfShortBlocks x L]
subWindows = current_frame.reshape(subWinLength,
numOfShortBlocks,
order='F').copy()
# Compute normalized sub-frame energies:
s = np.sum(subWindows**2, axis=0) / (Eol + eps)
# Compute entropy of the normalized sub-frame energies:
entropy = -np.sum(s * np.log2(s + eps))
self.energy_entropy.append(entropy)
self.energy_entropy = np.asarray(self.energy_entropy)