def createMelSpectrogram(input_path, fileName, output_path, saveOrShow=0):
# load sound signal
signal, sr = librosa.load(os.path.join(input_path, fileName), duration=10, sr=16000)
#signal = filter_signal(signal, sr, target_audio_length)
# create Mel Spectrogram
S = Melspectrogram(n_dft=1024,
n_hop=320,
#n_hop=256,
input_shape=(1, signal.shape[0]),
padding='same', sr=sr, n_mels=224, fmin=1400, fmax=sr/2,
power_melgram=2.0, return_decibel_melgram=True,
trainable_fb=False, trainable_kernel=False)(signal.reshape(1, 1, -1)).numpy()
S = S.reshape(S.shape[1], S.shape[2])
print(S.shape)
if saveOrShow == 0:
matplotlib.image.imsave(os.path.join(output_path, fileName.split(".")[0] + ".png"), S, cmap='inferno')
else:
#plt.imshow(S)
#plt.show()
display.specshow(S, sr=sr)
plt.show()
def createMelSpectrogramNew(input_path, fileName, output_path_train, output_path_val):
# load sound signal
signal, sr = librosa.load(os.path.join(input_path, fileName), sr=16000, mono=True)
abs_signal = [np.abs(s) for s in signal]
rolling_5s_abs_signal = [sum(abs_signal[i*5*sr:(i*5+5)*sr]) for i in range(int(len(abs_signal) // sr // 5))]
if len(signal) <= sr * 5:
# add 0 padding
signal = list(signal) + [0 for i in range(sr*5 - len(signal))]
signal = np.array(signal, dtype=np.float32)
# draw random number
rand = np.random.randint(0, 10)
if rand <= 7:
output_path = output_path_train
toAug = 1
else:
output_path = output_path_val
toAug = 0
if toAug == 0:
# normal mel spectrogram
S = Melspectrogram(n_dft=1024, n_hop=320, input_shape=(1, signal.shape[0]),
padding='same', sr=sr, n_mels=224, fmin=1400, fmax=sr/2,
power_melgram=2.0, return_decibel_melgram=True,
trainable_fb=False, trainable_kernel=False)(signal.reshape(1, 1, -1)).numpy()
S = S.reshape(S.shape[1], S.shape[2])
matplotlib.image.imsave(os.path.join(output_path, fileName.split(".")[0] + ".png"), S, cmap='inferno')
else:
# augmentation
mySignal = augmenter(signal, sr)
S = Melspectrogram(n_dft=1024, n_hop=320, input_shape=(1, signal.shape[0]),
padding='same', sr=sr, n_mels=224, fmin=1400, fmax=sr/2,
power_melgram=2.0, return_decibel_melgram=True,
trainable_fb=False, trainable_kernel=False)(mySignal.reshape(1, 1, -1)).numpy()
S = S.reshape(S.shape[1], S.shape[2])
matplotlib.image.imsave(os.path.join(output_path, fileName.split(".")[0] + "_noise.png"),
S, cmap='inferno')
else:
q_signal = np.quantile(rolling_5s_abs_signal, 0.75)
count = 0
numSamples = int((len(signal) // sr) // 5)
for i in range(numSamples):
tmpSignal = signal[int(i*5)*sr:int((i*5+5))*sr]
# cut out region is highest intensity
#window = 224 * 256
#interval = int((sr * 5 - window) // 10)
#intensity = [sum(tmpSignal[i*interval:i*interval+window]) for i in range(10)]
#idx = intensity.index(max(intensity))
#tmpSignal = tmpSignal[idx*interval:idx*interval+window]
#mask = [1 if np.abs(s) > median_signal else 0 for s in tmpSignal]
if sum([abs(j) for j in tmpSignal]) >= q_signal:
# draw random number
rand = np.random.randint(0, 10)
if rand <= 7:
output_path = output_path_train
toAug = 1
else:
output_path = output_path_val
toAug = 0
if toAug == 0:
S = Melspectrogram(n_dft=1024, n_hop=320, input_shape=(1, signal.shape[0]),
padding='same', sr=sr, n_mels=224, fmin=1400, fmax=sr/2,
power_melgram=2.0, return_decibel_melgram=True,
trainable_fb=False,
trainable_kernel=False)(tmpSignal.reshape(1, 1, -1)).numpy()
S = S.reshape(S.shape[1], S.shape[2])
matplotlib.image.imsave(os.path.join(output_path,
fileName.split(".")[0] + "_" + str(count) + ".png"),
S, cmap='inferno')
else:
# augmentation
mySignal = augmenter(tmpSignal, sr)
S = Melspectrogram(n_dft=1024, n_hop=320, input_shape=(1, signal.shape[0]),
padding='same', sr=sr, n_mels=224, fmin=1400, fmax=sr/2,
power_melgram=2.0, return_decibel_melgram=True,
trainable_fb=False,
trainable_kernel=False)(mySignal.reshape(1, 1, -1)).numpy()
S = S.reshape(S.shape[1], S.shape[2])
matplotlib.image.imsave(os.path.join(output_path,
fileName.split(".")[0] + "_" + str(count) + "_noise.png"),
S, cmap='inferno')
count += 1
d.extend(data_raw)
frames = (np.hstack(d)).flatten()
melspec = Melspectrogram(n_dft=1024,
n_hop=256,
input_shape=(1, frames.shape[0]),
padding='same',
sr=sr,
n_mels=224,
fmin=1400,
fmax=sr / 2,
power_melgram=2.0,
return_decibel_melgram=True,
trainable_fb=False,
trainable_kernel=False)(
frames.reshape(1, 1, -1)).numpy()
melspec = melspec.reshape(melspec.shape[1], melspec.shape[2])
print(
f"Frames array: {frames.shape}, Melspec array: {melspec.shape}"
)
melplot = display.specshow(melspec, sr=sr)
melplot.set_frame_on(False)
plt.tight_layout(pad=0)
plt.draw()
plt.pause(0.0001)
plt.clf()
if (melspec.shape[1] >= IM_SIZE[0]):
img = Image.frombuffer("RGBA",
fig.canvas.get_width_height(),
fig.canvas.buffer_rgba(), "raw",
"RGBA", 0, 1)
img = img.convert('RGB').resize(IM_SIZE[0:2])
