def __call__(self, noise, clean1, clean2, face1, face2):
noise = xp.asarray(noise).astype(xp.float32)
clean1 = xp.asarray(clean1).astype(xp.float32)
clean2 = xp.asarray(clean2).astype(xp.float32)
face1 = xp.asarray(face1).astype(xp.float32)[:, :, :, xp.newaxis]
face2 = xp.asarray(face2).astype(xp.float32)[:, :, :, xp.newaxis]
clean = xp.concatenate((clean1, clean2), axis=3)
compressed_noise, _ = op.compress_audio(noise)
compressed_clean, _ = op.compress_audio(clean)
mask1, mask2 = self.estimate_mask(spec=compressed_noise,
face1=face1,
face2=face2)
separated1 = op.mul(mask1, compressed_noise)
separated2 = op.mul(mask2, compressed_noise)
separated = F.concat((separated1, separated2),
axis=3) # (6, 2, 301, 514)
loss = evaluate_loss(self, separated, compressed_clean)
return loss
def __call__(self, noise, clean):
noise = xp.asarray(noise).astype(xp.float32)
clean = xp.asarray(clean).astype(xp.float32)
compressed_noise, _ = op.compress_audio(noise)
compressed_clean, _ = op.compress_audio(clean)
mask, _ = self.estimate_mask(spec=compressed_noise)
separated = op.mul(mask, compressed_noise)
loss = evaluate_loss(self, separated, compressed_clean)
return loss
def predict(model):
print("estimate mask...")
if env.INPUT_FACE == 0:
noise, clean1 = dataset.load_dataset_audio(
list([
env.TRAIN + 7, env.TRAIN + 1, env.TRAIN + 8, env.TRAIN + 3,
env.TRAIN + 4, env.TRAIN + 5, env.TRAIN + 6
]))
compressed_noise, _ = op.compress_audio(noise)
mask1, mask2 = model.estimate_mask(spec=compressed_noise)
elif env.INPUT_FACE == 1:
noise, clean1, face1 = dataset.load_dataset_single(
list([
env.TRAIN, env.TRAIN + 1, env.TRAIN + 2, env.TRAIN + 3,
env.TRAIN + 4, env.TRAIN + 5, env.TRAIN + 6
]))
compressed_noise, _ = op.compress_audio(noise)
mask1, mask2 = model.estimate_mask(spec=compressed_noise, face=face1)
else:
noise, clean1, clean2, face1, face2 = dataset.load_dataset_double(
list([
env.TRAIN, env.TRAIN + 1, env.TRAIN + 2, env.TRAIN + 3,
env.TRAIN + 4, env.TRAIN + 5, env.TRAIN + 6
]))
compressed_noise, _ = op.compress_audio(noise)
mask1, mask2 = model.estimate_mask(spec=compressed_noise,
face1=face1,
face2=face2)
print("mul mask...")
compressed_separated1 = op.mul(mask1, compressed_noise)
compressed_separated2 = op.mul(mask2, compressed_noise)
compressed_clean1, _ = op.compress_audio(clean1)
loss = evaluate_loss(model, compressed_separated1, compressed_clean1)
print(loss)
print("reconstruct audio...")
n = op.reconstruct_audio_complex(chainer.cuda.to_cpu(compressed_noise))
c1 = op.reconstruct_audio_complex(chainer.cuda.to_cpu(compressed_clean1))
# c2 = op.reconstruct_audio_complex(chainer.cuda.to_cpu(compressed_clean2))
y1 = op.reconstruct_audio_complex(
chainer.cuda.to_cpu(compressed_separated1.data))
y2 = op.reconstruct_audio_complex(
chainer.cuda.to_cpu(compressed_separated2.data))
print("save files...")
for i in range(n.shape[2]):
print("{0}/{1}".format(i + 1, n.shape[2]))
util.istft_and_save(
"{}/{}-synthesis.wav".format(os.environ['RESULT_DIR'], i), n[:, :,
i])
util.istft_and_save(
"{}/{}-clean1.wav".format(os.environ['RESULT_DIR'], i), c1[:, :,
i])
# util.istft_and_save("{}/{}-clean2.wav".format(os.environ['RESULT_DIR'], i), c2[:, :, i])
util.istft_and_save(
"{}/{}-separated1.wav".format(os.environ['RESULT_DIR'], i),
y1[:, :, i])
util.istft_and_save(
"{}/{}-separated2.wav".format(os.environ['RESULT_DIR'], i),
y2[:, :, i])