def gen_data(self,
dir_in,
dir_out=None,
file_id_list=None,
id_list=None,
return_dict=False):
"""
Prepare atom labels from wav files.
If id_list is not None, only the ids listed there are generated, otherwise for each .wav file in the dir_in.
Atoms are computed by the wcad algorithm. Examples with more than 70 atoms are rejected. One can create
a new file_id_list by uncommenting the lines before the return statement. Nevertheless, the current file_id_list
is not substituted by it. The algorithm also saves the extracted phrase component in dir_out/id_name.phrase,
if dir_out is not None.
:param dir_in: Directory containing the original wav files.
:param dir_out: Directory where the labels are stored. If None, no labels are stored.
:param file_id_list: Name of the file containing the ids. Normalisation parameters are saved using
this name to differentiate parameters between subsets.
:param id_list: The list of utterances to process.
Should have the form uttId1 \\n uttId2 \\n ...\\n uttIdN.
If None, all wav files in audio_dir are used.
:param return_dict: If True, returns an OrderedDict of all samples as first output.
:return: Returns mean=0.0, std_dev, min, max of atoms.
"""
# Fill file_id_list by .wav files in dir_in if not given and set an appropriate file_id_list_name.
if id_list is None:
id_list = list()
filenames = glob.glob(os.path.join(dir_in, "*.wav"))
for filename in filenames:
id_list.append(os.path.splitext(os.path.basename(filename))[0])
file_id_list_name = "all"
else:
file_id_list_name = os.path.splitext(
os.path.basename(file_id_list))[0]
if dir_out is not None:
makedirs_safe(dir_out)
if return_dict:
label_dict = OrderedDict()
mean_std_ext_atom = MeanStdDevExtractor()
min_max_ext_atom = MinMaxExtractor()
mean_std_ext_phrase = MeanStdDevExtractor()
min_max_ext_phrase = MinMaxExtractor()
# Compute Atoms.
from wcad import WaveInput, PitchExtractor, MultiphraseExtractor, DictionaryGenerator, AtomExtrator, ModelCreator, ModelSaver, Params, Paths
correct_utts = list()
self.logger.info("Create atom labels for " +
"[{0}]".format(", ".join(str(i) for i in id_list)))
for id_name in id_list:
self.logger.debug("Create atom labels for " + id_name)
# Wcad has to be called in its root directory, therefore a change dir operation is necessary.
cwd = os.getcwd()
os.chdir(self.wcad_root)
args = [dir_in + "/" + id_name + ".wav", dir_out]
print(args)
params = Params()
# Overwrite the possible theta values by selected values.
params.local_atoms_thetas = self.theta_interval
params.k = [self.k]
# params.min_atom_amp = 0.1
paths = Paths(args, params)
# Start the extraction process.
start_t = time.time()
waveform = WaveInput(paths.wav, params).read()
pitch = PitchExtractor(waveform, params, paths).compute()
# Compute the phrase component.
phrase = MultiphraseExtractor(pitch, waveform, params,
paths).compute()
phrase_curve = phrase.curve
# Extract atroms.
dictionary = DictionaryGenerator(params, paths).compute()
atoms = AtomExtrator(waveform, pitch, phrase, dictionary, params,
paths).compute()
# Create a model.
model = ModelCreator(phrase, atoms, pitch).compute()
print(('Model created in %s seconds' % (time.time() - start_t)))
# Save the atoms.
ModelSaver(model, params, paths).save()
os.chdir(cwd)
# Check if output can be correct.
possible_extraction_failure = False
if len(atoms) < 50 and not any(a.amp > 10 for a in atoms):
correct_utts.append(id_name)
else:
self.logger.warning("Possible fail of atom extractor for " +
id_name + " (atoms: " + str(len(atoms)) +
", frames: " + str(len(phrase_curve)) +
", max: " +
str(max(a.amp for a in atoms)) + ").")
possible_extraction_failure = True
atoms.sort(key=lambda x: x.position)
# print_atoms(atoms)
# Get audio length needed to trim the atoms.
duration = self.get_audio_length(id_name, dir_in,
self.frame_size_ms)
# The algorithm generates a few atoms at negative positions,
# pad them into the first atom at positive position.
padded_amp = 0
padded_theta = 0
for idx, atom in enumerate(atoms):
if atom.position < 0:
padded_amp += atom.amp
padded_theta += atom.theta
else:
atoms[idx].amp += padded_amp # Pad the amplitude.
atoms[idx].theta = (atoms[idx].theta +
padded_theta) / (idx + 1)
del atoms[:idx] # Remove the negative atoms from the list.
break
# print_atoms(atoms)
# The algorithm might also generate a few atoms beyond the last label,
# pad them into the last label.
padded_amp = 0
padded_theta = 0
for idx, atom in reversed(list(enumerate(atoms))):
if atom.position * self.frame_size_ms > duration:
padded_amp += atom.amp
padded_theta += atom.theta
else:
atoms[idx].amp += padded_amp
atoms[idx].theta = (atoms[idx].theta +
padded_theta) / (len(atoms) - idx)
atoms = atoms[:-(len(atoms) - idx - 1)
or None] # Remove atoms beyond last label.
break
# print_atoms(atoms)
# Create a label for each frame (size of frame_size_ms) with amplitude and theta of contained atoms.
np_atom_labels = AtomLabelGen.atoms_to_labels(
atoms, self.theta_interval, int(duration / self.frame_size_ms))
np_atom_amps = np.sum(np_atom_labels, axis=1)
if not possible_extraction_failure: # Only add successful extractions to mean and std_dev computation.
mean_std_ext_atom.add_sample(
np_atom_amps[np_atom_amps[:, 0] != 0.0]
) # Only compute std_dev from atoms.
min_max_ext_atom.add_sample(np_atom_amps)
# mean_std_ext_phrase.add_sample(phrase_curve)
# min_max_ext_phrase.add_sample(phrase_curve)
if return_dict:
label_dict[id_name] = np_atom_labels
if dir_out is not None:
# Save phrase, because it might be used in synthesis.
phrase_curve.astype('float32').tofile(
os.path.join(dir_out, id_name + self.ext_phrase))
# Save atoms binary (float32).
np_atom_labels.astype('float32').tofile(
os.path.join(dir_out, id_name + self.ext_atoms))
# Create a readable version of the atom data.
# np.savetxt(os.path.join(dir_out, id_name + self.ext_atoms + ".txt"), np_atom_labels)
# Manually set mean of atoms to 0, otherwise frames without atom will have an amplitude.
mean_std_ext_atom.sum_frames[:] = 0.0
mean_std_ext_atom.sum_squared_frames[
1] = mean_std_ext_atom.sum_length * self.theta_interval[-1]
mean_std_ext_atom.save(os.path.join(dir_out, file_id_list_name))
min_max_ext_atom.save(os.path.join(dir_out, file_id_list_name))
# mean_std_ext_phrase.save(os.path.join(dir_out, file_id_list_name + '-phrase'))
# min_max_ext_phrase.save(os.path.join(dir_out, file_id_list_name + '-phrase'))
mean_atoms, std_atoms = mean_std_ext_atom.get_params()
min_atoms, max_atoms = min_max_ext_atom.get_params()
# mean_phrase, std_phrase = mean_std_ext_phrase.get_params()
# min_phrase, max_phrase = min_max_ext_atom.get_params()
# Use this block to save the part of the file_id_list for which atom extraction was successful into a new file.
if correct_utts:
with open(
os.path.join(
os.path.dirname(dir_in), "wcad_" +
os.path.basename(file_id_list_name) + ".txt"),
'w') as f:
f.write('\n'.join(correct_utts) + '\n')
if return_dict:
# Return dict of labels for all utterances.
return label_dict, \
mean_atoms, std_atoms, \
min_atoms, max_atoms
# mean_phrase, std_phrase, \
# min_phrase, max_phrase
else:
return mean_atoms, std_atoms, \
min_atoms, max_atoms
def gen_data(self, dir_in, dir_out=None, file_id_list=None, id_list=None, add_deltas=False, return_dict=False):
"""
Prepare LF0 and V/UV features from audio files. If add_delta is false each numpy array has the dimension
num_frames x 2 [f0, vuv], otherwise the deltas and double deltas are added between
the features resulting in num_frames x 4 [lf0(3*1), vuv].
:param dir_in: Directory where the .wav files are stored for each utterance to process.
:param dir_out: Main directory where the labels and normalisation parameters are saved to subdirectories.
If None, labels are not saved.
:param file_id_list: Name of the file containing the ids. Normalisation parameters are saved using
this name to differentiate parameters between subsets.
:param id_list: The list of utterances to process.
Should have the form uttId1 \\n uttId2 \\n ...\\n uttIdN.
If None, all file in audio_dir are used.
:param add_deltas: Add deltas and double deltas to all features except vuv.
:param return_dict: If true, returns an OrderedDict of all samples as first output.
:return: Returns two normalisation parameters as tuple. If return_dict is True it returns
all processed labels in an OrderedDict followed by the two normalisation parameters.
"""
# Fill file_id_list by .wav files in dir_in if not given and set an appropriate file_id_list_name.
if id_list is None:
id_list = list()
filenames = glob.glob(os.path.join(dir_in, "*.wav"))
for filename in filenames:
id_list.append(os.path.splitext(os.path.basename(filename))[0])
file_id_list_name = "all"
else:
file_id_list_name = os.path.splitext(os.path.basename(file_id_list))[0]
# Create directories in dir_out if it is given.
if dir_out is not None:
if add_deltas:
makedirs_safe(os.path.join(dir_out, LF0LabelGen.dir_deltas))
else:
makedirs_safe(os.path.join(dir_out, LF0LabelGen.dir_lf0))
makedirs_safe(os.path.join(dir_out, LF0LabelGen.dir_vuv))
# Create the return dictionary if required.
if return_dict:
label_dict = OrderedDict()
# Create normalisation computation units.
norm_params_ext_lf0 = MeanStdDevExtractor()
# norm_params_ext_vuv = MeanStdDevExtractor()
norm_params_ext_deltas = MeanStdDevExtractor()
logging.info("Extract WORLD LF0 features for " + "[{0}]".format(", ".join(str(i) for i in id_list)))
for file_name in id_list:
logging.debug("Extract WORLD LF0 features from " + file_name)
# Load audio file and extract features.
audio_name = os.path.join(dir_in, file_name + ".wav")
raw, fs = soundfile.read(audio_name)
_f0, t = pyworld.dio(raw, fs) # Raw pitch extraction. TODO: Use magphase here?
f0 = pyworld.stonemask(raw, _f0, t, fs) # Pitch refinement.
# Compute lf0 and vuv information.
lf0 = np.log(f0, dtype=np.float32)
lf0[lf0 <= math.log(LF0LabelGen.f0_silence_threshold)] = LF0LabelGen.lf0_zero
lf0, vuv = interpolate_lin(lf0)
if add_deltas:
# Compute the deltas and double deltas for all features.
lf0_deltas, lf0_double_deltas = compute_deltas(lf0)
# Combine them to a single feature sample.
labels = np.concatenate((lf0, lf0_deltas, lf0_double_deltas, vuv), axis=1)
# Save into return dictionary and/or file.
if return_dict:
label_dict[file_name] = labels
if dir_out is not None:
labels.tofile(os.path.join(dir_out, LF0LabelGen.dir_deltas, file_name + LF0LabelGen.ext_deltas))
# Add sample to normalisation computation unit.
norm_params_ext_deltas.add_sample(labels)
else:
# Save into return dictionary and/or file.
if return_dict:
label_dict[file_name] = np.concatenate((lf0, vuv), axis=1)
if dir_out is not None:
lf0.tofile(os.path.join(dir_out, LF0LabelGen.dir_lf0, file_name + LF0LabelGen.ext_lf0))
vuv.astype(np.float32).tofile(os.path.join(dir_out, LF0LabelGen.dir_vuv, file_name + LF0LabelGen.ext_vuv))
# Add sample to normalisation computation unit.
norm_params_ext_lf0.add_sample(lf0)
# norm_params_ext_vuv.add_sample(vuv)
# Save mean and std dev of all features.
if not add_deltas:
norm_params_ext_lf0.save(os.path.join(dir_out, LF0LabelGen.dir_lf0, file_id_list_name))
# norm_params_ext_vuv.save(os.path.join(dir_out, LF0LabelGen.dir_vuv, file_id_list_name))
else:
# Manually set vuv normalisation parameters before saving.
norm_params_ext_deltas.sum_frames[-1] = 0.0 # Mean = 0.0
norm_params_ext_deltas.sum_squared_frames[-1] = norm_params_ext_deltas.sum_length # Variance = 1.0
norm_params_ext_deltas.save(os.path.join(dir_out, LF0LabelGen.dir_deltas, file_id_list_name))
# Get normalisation parameters.
if not add_deltas:
norm_lf0 = norm_params_ext_lf0.get_params()
# norm_vuv = norm_params_ext_vuv.get_params()
norm_first = np.concatenate((norm_lf0[0], (0.0,)), axis=0)
norm_second = np.concatenate((norm_lf0[1], (1.0,)), axis=0)
else:
norm_first, norm_second = norm_params_ext_deltas.get_params()
if return_dict:
# Return dict of labels for all utterances.
return label_dict, norm_first, norm_second
else:
return norm_first, norm_second