def seperate(self, spectra, cmvn=None, apply_log=True):
"""
spectra: stft complex results T x F
cmvn: python dict contains global mean/std
apply_log: using log-spectrogram or not
"""
if not np.iscomplexobj(spectra):
raise ValueError("Input must be matrix in complex value")
input_spectra = np.log(np.maximum(
np.abs(spectra), EPSILON)) if apply_log else np.abs(spectra)
input_spectra = apply_cmvn(input_spectra,
cmvn) if cmvn else input_spectra
mix_spec, spk1_spec, spk2_spec, speaker_1, speaker_2, Orth_const = self.nnet(
th.tensor(input_spectra, dtype=th.float32, device=self.location),
None,
per_train=False)
mask = []
mask.append(spk1_spec /
(spk1_spec + spk2_spec + th.finfo(th.float32).eps))
mask.append(spk2_spec /
(spk1_spec + spk2_spec + th.finfo(th.float32).eps))
return mask, [
spectra * spk_mask[0].cpu().data.numpy() for spk_mask in mask
]
def _transform(self, mixture_specs, targets_specs_list):
"""
Transform original spectrogram
If mixture_specs is a complex object, it means PAM will be used for training
It can be configured in .yaml, egs: apply_abs=false to produce complex results
If mixture_specs is real, we will using AM(ratio mask)
Arguments:
mixture_specs: non-log complex/real spectrogram
targets_specs_list: list of non-log complex/real spectrogram for each target speakers
Returns:
python dictionary with four attributes:
num_frames: length of current utterance
feature: input feature for networks, egs: log spectrogram + cmvn
source_attr: a dictionary with at most 2 keys: spectrogram and phase(for PSM), each contains a tensor
target_attr: same keys like source_attr, each keys correspond to a tensor list
"""
# apply_log and cmvn, for nnet input
# NOTE: mixture_specs may be complex or real
log_spectra = np.log(
np.maximum(
np.abs(mixture_specs)
if np.iscomplexobj(mixture_specs) else mixture_specs, EPSILON))
if self.mvn_dict:
log_spectra = apply_cmvn(log_spectra, self.mvn_dict)
# using dict to pack infomation needed in loss
source_attr = {}
target_attr = {}
if np.iscomplexobj(mixture_specs):
source_attr["spectrogram"] = th.tensor(np.abs(mixture_specs),
dtype=th.float32)
target_attr["spectrogram"] = [
th.tensor(np.abs(t), dtype=th.float32)
for t in targets_specs_list
]
source_attr["phase"] = th.tensor(np.angle(mixture_specs),
dtype=th.float32)
target_attr["phase"] = [
th.tensor(np.angle(t), dtype=th.float32)
for t in targets_specs_list
]
else:
source_attr["spectrogram"] = th.tensor(mixture_specs,
dtype=th.float32)
target_attr["spectrogram"] = [
th.tensor(t, dtype=th.float32) for t in targets_specs_list
]
return {
"num_frames": mixture_specs.shape[0],
"feature": th.tensor(log_spectra, dtype=th.float32),
"source_attr": source_attr,
"target_attr": target_attr
}
def seperate(self, spectra, cmvn=None):
"""
spectra: stft complex results T x F
cmvn: python dict contains global mean/std
"""
if not np.iscomplexobj(spectra):
raise ValueError("Input must be matrix in complex value")
# compute log-magnitude spectrogram
log_spectra = np.log(np.maximum(np.abs(spectra), EPSILON))
# apply cmvn or not
log_spectra = apply_cmvn(log_spectra, cmvn) if cmvn else log_spectra
out_masks = self.nnet(th.tensor(log_spectra,
dtype=th.float32,
device=self.location),
train=False)
spk_masks = [spk_mask.cpu().data.numpy() for spk_mask in out_masks]
return spk_masks, [spectra * spk_mask for spk_mask in spk_masks]
def _transform(self, mixture_specs, targets_specs_list):
"""
Transform from numpy/list to torch types
"""
# compute vad mask before cmvn
vad_mask = compute_vad_mask(
mixture_specs, self.vad_threshold, apply_exp=True)
# apply cmvn
if self.mvn_dict:
mixture_specs = apply_cmvn(mixture_specs, self.mvn_dict)
# compute target embedding index
target_attr = np.argmax(np.array(targets_specs_list), 0)
return {
"num_frames": mixture_specs.shape[0],
"spectrogram": th.tensor(mixture_specs, dtype=th.float32),
"target_attr": th.tensor(target_attr, dtype=th.int64),
"silent_mask": th.tensor(vad_mask, dtype=th.float32)
}
def _transform(self, mixture_specs, targets_specs_list):
"""
Transform from numpy/list to torch types
"""
# compute vad mask before cmvn
vad_mask = compute_vad_mask(mixture_specs,
self.vad_threshold,
apply_exp=True)
# apply cmvn
if self.mvn_dict:
mixture_specs = apply_cmvn(mixture_specs, self.mvn_dict)
# compute target embedding index
target_attr = np.argmax(np.array(targets_specs_list), 0)
return {
"num_frames": mixture_specs.shape[0],
"spectrogram": th.tensor(mixture_specs, dtype=th.float32),
"target_attr": th.tensor(target_attr, dtype=th.int64),
"silent_mask": th.tensor(vad_mask, dtype=th.float32)
}
def seperate(self, spectra, cmvn=None):
"""
spectra: stft complex results T x F
cmvn: python dict contains global mean/std
"""
if not np.iscomplexobj(spectra):
raise ValueError("Input must be matrix in complex value")
# compute log-magnitude spectrogram
log_spectra = np.log(np.maximum(np.abs(spectra), EPSILON))
# compute vad mask before do mvn
vad_mask = compute_vad_mask(
log_spectra, threshold_db=40).astype(np.bool)
# print("Keep {} bins out of {}".format(np.sum(vad_mask), vad_mask.size))
pca_mat, spk_masks = self._cluster(
apply_cmvn(log_spectra, cmvn) if cmvn else log_spectra, vad_mask)
return pca_mat, spk_masks, [
spectra * spk_mask for spk_mask in spk_masks
]
def seperate(self, spectra, cmvn=None):
"""
spectra: stft complex results T x F
cmvn: python dict contains global mean/std
"""
if not np.iscomplexobj(spectra):
raise ValueError("Input must be matrix in complex value")
# compute log-magnitude spectrogram
log_spectra = np.log(np.maximum(np.abs(spectra), EPSILON))
# compute vad mask before do mvn
vad_mask = compute_vad_mask(log_spectra,
threshold_db=40).astype(np.bool)
# print("Keep {} bins out of {}".format(np.sum(vad_mask), vad_mask.size))
pca_mat, spk_masks = self._cluster(
apply_cmvn(log_spectra, cmvn) if cmvn else log_spectra, vad_mask)
return pca_mat, spk_masks, [
spectra * spk_mask for spk_mask in spk_masks
]
