def compute_mask(mixture, targets_list, mask_type):
"""
Arguments:
mixture: STFT of mixture signal(complex result)
targets_list: python list of target signal's STFT results(complex result)
mask_type: ["irm", "ibm", "iam", "psm"]
Return:
masks_list
"""
if mask_type == "ibm":
max_index = np.argmax(
np.stack([cmat_abs(mat) for mat in targets_list]), 0)
return [max_index == s for s in range(len(targets_list))]
if mask_type == "irm":
denominator = sum([cmat_abs(mat) for mat in targets_list]) + EPSILON
else:
denominator = cmat_abs(mixture) + EPSILON
if mask_type != "psm":
masks = [cmat_abs(mat) / denominator for mat in targets_list]
else:
mixture_phase = np.angle(mixture)
masks = [
cmat_abs(mat) * np.cos(mixture_phase - np.angle(mat)) / denominator
for mat in targets_list
]
return masks
def compute_mask(speech, noise_or_mixture, mask):
"""
for signal model:
y = x1 + x2
def f = STFT(x):
f(y) = f(x1) + f(x2) => |f(y)| = |f(x1) + f(x2)| < |f(x1)| + |f(x2)|
for irm:
1) M(x1) = |f(x1)| / (|f(x1)| + |f(x2)|) DongYu
2) M(x1) = |f(x1)| / sqrt(|f(x1)|^2 + |f(x2)|^2) DeliangWang
s.t. 1 >= 2) >= 1) >= 0
for iam(FFT-mask, smm):
M(x1) = |f(x1)| / |f(y)| = |f(x1)| / |f(x1) + f(x2)| in [0, oo]
for psm:
M(x1) = |f(x1) / f(y)| = |f(x1)| * cos(delta_phase) / |f(y)|
"""
if mask == "ibm":
binary_mask = cmat_abs(speech) > cmat_abs(noise_or_mixture)
return binary_mask.astype(np.float)
# irm/iam/psm
if mask == "irm":
# denominator = cmat_abs(speech) + cmat_abs(noise_or_mixture)
denominator = np.sqrt(
cmat_abs(speech)**2 + cmat_abs(noise_or_mixture)**2)
else:
denominator = cmat_abs(noise_or_mixture)
if mask == "psm":
return cmat_abs(speech) * np.cos(
np.angle(noise_or_mixture) - np.angle(speech)) / denominator
elif mask == "psa":
# keep nominator only
return cmat_abs(speech) * np.cos(
np.angle(noise_or_mixture) - np.angle(speech))
else:
# irm/iam
return cmat_abs(speech) / denominator
def compute_mask(tgt, mix, mask):
"""
for signal model:
y = x1 + x2
def f = STFT(x):
f(y) = f(x1) + f(x2) => |f(y)| = |f(x1) + f(x2)| < |f(x1)| + |f(x2)|
for irm:
1) M(x1) = |f(x1)| / (|f(x1)| + |f(x2)|) DongYu
2) M(x1) = |f(x1)| / sqrt(|f(x1)|^2 + |f(x2)|^2) Deliang Wang
s.t. 1 >= 2) >= 1) >= 0
for iam(FFT-mask, smm):
M(x1) = |f(x1)| / |f(y)| = |f(x1)| / |f(x1) + f(x2)| in [0, oo]
for psm:
M(x1) = |f(x1) / f(y)| = |f(x1)| * cos(delta_phase) / |f(y)|
for crm:
M(x1) = f(x1) / f(y)
"""
# target speech
tgt_abs = cmat_abs(tgt)
# mixture
mix_abs = cmat_abs(mix)
# interference speech
inf_abs = cmat_abs(mix - tgt)
if mask == "ibm":
return (tgt_abs > inf_abs).astype(np.float32)
# irm/iam/psm
if mask == "irm":
# denominator = tgt_abs + inf_abs
denominator = np.sqrt(tgt_abs**2 + inf_abs**2 + EPSILON)
elif mask == "crm":
denominator = mix + EPSILON
else:
denominator = mix_abs
if mask == "psm":
return tgt_abs * np.cos(np.angle(mix) - np.angle(tgt)) / denominator
# phase sensitive amplitude
elif mask == "psa":
# keep nominator only
non_neg = np.maximum(0, np.cos(np.angle(mix) - np.angle(tgt)))
return tgt_abs * non_neg
elif mask == "crm":
# stack real/imag part
cpx_mask = tgt / denominator
return np.hstack(
[tangent(np.real(cpx_mask)),
tangent(np.imag(cpx_mask))])
else:
# irm/iam
return tgt_abs / denominator
def compute_vad_masks(spectrogram, proportion):
"""
We ignore several minimum values and keep proportion*100% energy
Arguments:
spectrogram: F x T
Return:
vad_mask: T x F
"""
energy_mat = cmat_abs(spectrogram)
energy_vec = np.sort(energy_mat.flatten())
filter_energy = np.sum(energy_vec) * (1 - proportion)
threshold = 0
cumsum, index = 0, 0
while index < energy_vec.shape[0]:
threshold = energy_vec[index]
cumsum += threshold
if cumsum > filter_energy:
break
index += 1
# silence if 1
vad_mask = (energy_mat < threshold)
return vad_mask.transpose(), index