def parse_audio(self, audio_path):
if self.aug_conf and self.aug_conf.speed_volume_perturb:
y = load_randomly_augmented_audio(audio_path, self.sample_rate)
else:
y = load_audio(audio_path)
if self.noise_injector:
add_noise = np.random.binomial(1, self.aug_conf.noise_prob)
if add_noise:
y = self.noise_injector.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)
# 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)
if self.aug_conf and self.aug_conf.spec_augment:
spect = spec_augment(spect)
return spect
def parse_audio(self, audio_path):
if self.aug_conf and self.aug_conf.speed_volume_perturb:
y = load_randomly_augmented_audio(audio_path, self.sample_rate)
else:
y = load_audio(audio_path)
if self.noise_injector:
add_noise = np.random.binomial(1, self.aug_conf.noise_prob)
if add_noise:
y = self.noise_injector.inject_noise(y)
##get ten file de ve hinh
# nanlist=audio_path.split("/")
# nanLs = nanlist[len(nanlist)-1]
# name = nanLs.split(".")[0]+"_"+nanLs.split(".")[1]
#tín hiệu thô
# fig1,ax22= plt.subplots()
# plt.title('Tín hiệu thô của câu nói \'anh có thể gọi cho tôi không\'')
# plt.plot(y)
# plt.xlabel('Sample')
# plt.ylabel('Amplitude')
# fig1.savefig('/work/Source/deepspeech.pytorch/deepspeech_pytorch/quyenImg/'+name+'tinhieutho'+'.png')
n_fft = int(self.sample_rate * self.window_size) #320
win_length = n_fft #320
hop_length = int(self.sample_rate * self.window_stride) #160
# STFT
D = librosa.stft(
y,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=self.window
) #array([[ 2.42148260e-01+0.00000000e+00j, -1.00020550e-01+0.00000000e+00j,
spect, phase = librosa.magphase(
D
) # S = log(S+1)##array([[ 2.42148260e-01+0.00000000e+00j, -1.00020550e-01+0.00000000e+00j,
# print("++**",audio_path)
# fig, ax = plt.subplots()
# img = librosa.display.specshow(librosa.amplitude_to_db(spect,ref=np.max), y_axis='log', x_axis='time', ax=ax)
# ax.set_title(audio_path)
# fig.savefig('/work/Source/deepspeech.pytorch/deepspeech_pytorch/quyenImg/'+name+'.png')
#có thanh độ lớn biên độ
# log_spectrogram = librosa.amplitude_to_db(spect)
# plt.figure(figsize=(12,8))
# librosa.display.specshow(log_spectrogram, sr=self.sample_rate,
# y_axis='log', x_axis='time',hop_length=160)
# plt.xlabel("Time")
# plt.ylabel("Frequency")
# plt.colorbar(format="%+2.0f dB")
# plt.title("Spectrogram (dB)")
# plt.savefig('/work/Source/deepspeech.pytorch/deepspeech_pytorch/quyenImg/'+name+'.png')
spect = np.log1p(
spect
) #tensor([[2.1684e-01, 9.5329e-02, 1.0469e-01, ..., 1.2308e-03, 2.3625e-03,
# fig2, ax2 = plt.subplots()
# img2 = librosa.display.specshow(librosa.amplitude_to_db(spect,ref=np.max), y_axis='log', x_axis='time', ax=ax)
# ax2.set_title(audio_path+"(log)")
# fig2.savefig('/work/Source/deepspeech.pytorch/deepspeech_pytorch/quyenImg/'+name+"(log)"+'.png')
spect = torch.FloatTensor(
spect
) #tensor([[2.1684e-01, 9.5329e-02, 1.0469e-01, ..., 1.2308e-03, 2.3625e-03,
if self.normalize:
mean = spect.mean() #tính trung bình cộng
#mean=np.log1p(mean)
std = spect.std()
#std=np.log1p(std)#độ lệch chuẩn
if (mean == torch.tensor(0) or std == torch.tensor(0)):
print("nan nan")
spect.add_(-mean)
spect.div_(std)
if self.aug_conf and self.aug_conf.spec_augment:
spect = spec_augment(spect)
return spect