def parse_audio(self, audio_path):
if self.augment:
y = load_randomly_augmented_audio(audio_path, self.sample_rate)
else:
y = load_audio(audio_path)
if self.noiseInjector:
logging.info("inject noise")
add_noise = np.random.binomial(1, self.noise_prob)
if add_noise:
y = self.noiseInjector.inject_noise(y)
n_fft = int(self.sample_rate * self.window_size)
win_length = n_fft
hop_length = int(self.sample_rate * self.window_stride)
# Short-time Fourier transform (STFT)
D = librosa.stft(y, n_fft=n_fft, hop_length=hop_length,
win_length=win_length, window=self.window)
spect, phase = librosa.magphase(D)
# S = log(S+1)
spect = np.log1p(spect)
spect = torch.FloatTensor(spect)
if self.normalize:
mean = spect.mean()
std = spect.std()
spect.add_(-mean)
spect.div_(std)
return spect
def transform_mfcc_from_file(fname, endian='in16', sr=16000, L=None, win_size=0.025, win_step=0.01, \
num_cep=13, nfilt=26, preemph=0.97, appendEnergy=True):
sr, buff = load_audio(fname, sr=sr, endian=endian)
if L:
buff = chop_audio(buff, L)
feat = transform_mfcc(buff, sr, win_size, win_step, num_cep, nfilt,
preemph, appendEnergy)
return feat