def process_dir(festival_dir, txt_dir, id_list, out_dir, custom_voice=None):
"""Create Utterance structures for all sentences in `id_list` and save them to `out_dir`.
Args:
festival_dir (str): Directory containing festival installation.
txt_dir (str): Directory containing text transcriptions.
id_list (str): List of file basenames to process.
out_dir (str): Directory to save the output to.
"""
file_ids = utils.get_file_ids(id_list=id_list)
sentences = []
# For all file_ids load the sentence and add a command to create and save the Utterance structure.
for file_id in sorted(file_ids):
sentence = file_io.load_lines(os.path.join(txt_dir,
f'{file_id}.txt'))[0]
sentence = sentence.replace('"', '\\"')
sentences.append(sentence)
# If the file_ids are paths (e.g. for multi-speaker data), make sure the directory structure is already in place.
utils.make_dirs(os.path.join(out_dir, 'utts'), file_ids)
# Create and save the Utterance structures.
create_utterances(festival_dir,
file_ids,
sentences,
out_dir,
custom_voice=custom_voice)
def plot_repeated_batch_f0(features, *repeated_predictions, out_dir=None):
if out_dir is not None:
plots_dir = os.path.join(out_dir, 'plots', 'repeated_f0')
make_dirs(plots_dir, features['name'])
n_frames = features['n_frames'].cpu().detach().numpy()
target_f0, target_vuv = utils.detach_batched_seqs(torch.exp(
features['lf0']),
features['vuv'],
seq_len=n_frames)
repeated_pred_f0 = utils.detach_batched_seqs(
*(torch.exp(predicted['lf0']) for predicted in repeated_predictions),
seq_len=n_frames)
for i, name in enumerate(features['name']):
if out_dir is None:
out_file = None
else:
out_file = os.path.join(plots_dir, f'{name}.pdf')
plot_repeated_f0(target_f0[i],
target_vuv[i],
*(pred_f0[i] for pred_f0 in repeated_pred_f0),
name=name,
out_file=out_file)
def process(wav_dir, id_list, out_dir, calculate_normalisation,
normalisation_of_deltas):
"""Processes wav files in id_list, saves the log-F0 and MVN parameters to files.
Args:
wav_dir (str): Directory containing the wav files.
id_list (str): List of file basenames to process.
out_dir (str): Directory to save the output to.
calculate_normalisation (bool): Whether to automatically calculate MVN parameters after extracting F0.
normalisation_of_deltas (bool): Also calculate the MVN parameters for the delta and delta delta features.
"""
file_ids = get_file_ids(wav_dir, id_list)
make_dirs(os.path.join(out_dir, 'lf0'), file_ids)
make_dirs(os.path.join(out_dir, 'vuv'), file_ids)
for file_id in file_ids:
wav_path = os.path.join(wav_dir, f'{file_id}.wav')
wav, sample_rate = file_io.load_wav(wav_path)
f0, vuv = analysis(wav, sample_rate)
lf0 = np.log(f0)
file_io.save_bin(lf0, os.path.join(out_dir, 'lf0', f'{file_id}.npy'))
file_io.save_bin(vuv, os.path.join(out_dir, 'vuv', f'{file_id}.npy'))
if calculate_normalisation:
process_mvn(out_dir,
'lf0',
id_list=id_list,
deltas=normalisation_of_deltas)
def create_utterances(festival_dir,
file_ids,
sentences,
out_dir,
custom_voice=None):
festival_exe = os.path.join(festival_dir, 'bin', 'festival')
scm_commands = [f'#!{festival_exe}']
if custom_voice is not None:
# Run Festival with a particular voice.
scm_commands.append(f'(voice_{custom_voice})')
scm_command_str = '(utt.save (utt.synth (Utterance Text "{sentence}" )) "{utt_file}")'
for file_id, sentence in zip(file_ids, sentences):
utt_file = os.path.join(out_dir, 'utts', f'{file_id}.utt')
scm_commands.append(
scm_command_str.format(sentence=sentence, utt_file=utt_file))
# Save the commands.
gen_utts_scm_file = os.path.join(out_dir, 'gen_utts.scm')
file_io.save_lines(scm_commands, gen_utts_scm_file)
# If the file_ids are paths (e.g. for multi-speaker data), make sure the directory structure is already in place.
utils.make_dirs(os.path.join(out_dir, 'utts'), file_ids)
# Run the commands.
scm_file = os.path.join(out_dir, 'gen_utts.scm')
# Argument `check=True` ensures that an exception is raised if the process' return code is non-zero.
subprocess.run([festival_exe, '-b', scm_file], check=True)
def process(lab_dir, id_list, out_dir, state_level):
"""Processes label files in id_list, saves the phone identities (as a string) to text files.
Args:
lab_dir (str): Directory containing the label files.
id_list (str): List of file basenames to process.
out_dir (str): Directory to save the output to.
state_level (bool): Indicates that the label files are state level if True, otherwise they are frame level.
"""
file_ids = utils.get_file_ids(id_list=id_list)
utils.make_dirs(os.path.join(out_dir, 'phones'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'n_phones'), file_ids)
for file_id in file_ids:
# Label processing.
lab_path = os.path.join(lab_dir, f'{file_id}.lab')
label = lab_to_feat.Label(lab_path, state_level)
phones = label.phones
n_phones = len(label.phones)
file_io.save_lines(phones,
os.path.join(out_dir, 'phones', f'{file_id}.txt'))
file_io.save_txt(n_phones,
os.path.join(out_dir, 'n_phones', f'{file_id}.txt'))
def plot_batch_f0(features, predicted, use_vuv=None, out_dir=None):
if use_vuv is None:
use_vuv = 'vuv' in predicted
if out_dir is not None:
plots_dir = os.path.join(out_dir, 'plots', 'f0')
make_dirs(plots_dir, features['name'])
n_frames = features['n_frames'].cpu().detach().numpy()
target_f0, target_vuv = utils.detach_batched_seqs(torch.exp(
features['lf0']),
features['vuv'],
seq_len=n_frames)
pred_f0 = utils.detach_batched_seqs(torch.exp(predicted['lf0']),
seq_len=n_frames)
if use_vuv:
pred_vuv = utils.detach_batched_seqs(predicted['vuv'] > 0.5,
seq_len=n_frames)
else:
pred_vuv = [None] * len(pred_f0)
for i, name in enumerate(features['name']):
if out_dir is None:
out_file = None
else:
out_file = os.path.join(plots_dir, f'{name}.pdf')
plot_f0(target_f0[i],
target_vuv[i],
pred_f0[i],
pred_vuv[i],
name=name,
out_file=out_file)
def save_dir(save_fn, path, data, file_ids, feat_ext=None):
utils.make_dirs(path, file_ids)
for datum, file_id in zip(data, file_ids):
if feat_ext is not None:
file_id = f'{file_id}.{feat_ext}'
file_path = os.path.join(path, file_id)
save_fn(datum, file_path)
def cluster(embeddings, n_clusters, names=None, out_dir=None):
"""Processes wav files in id_list, saves the log-F0 and MVN parameters to files.
Args:
embeddings_dir (str): Directory containing the embedding files.
n_clusters (int): Number of clusters for k-means.
names (str): List of file basenames to process.
out_dir (str): Directory to save the output to.
"""
if out_dir is not None:
if names is None:
raise ValueError(
'If `out_dir` is given, then `names` of individual sentences must also be given'
)
centres_path = os.path.join(out_dir, 'k_means', 'clusters')
make_dirs(centres_path, names)
assignments_path = os.path.join(out_dir, 'k_means',
'cluster_assignments')
make_dirs(assignments_path, names)
# Cluster with k-means.
kmeans = KMeans(n_clusters=n_clusters).fit(embeddings)
cluster_centres = kmeans.cluster_centers_
cluster_assignments = kmeans.labels_
# Save the cluster assignments and clusters to files.
if out_dir is not None:
cluster_names = [f'cluster_{i}' for i in range(n_clusters)]
file_io.save_dir(file_io.save_bin,
centres_path,
cluster_centres,
cluster_names,
feat_ext='npy')
file_io.save_dir(file_io.save_txt,
assignments_path,
cluster_assignments,
names,
feat_ext='txt')
counts = np.array([(i,
cluster_assignments.reshape(-1).tolist().count(i))
for i in range(n_clusters)])
file_io.save_txt(counts, f'{assignments_path}_counts.txt')
return cluster_centres, cluster_assignments
def _add_alignments_to_lab(self, mlf, lab_align_dir, lab_dir, file_ids):
make_dirs(lab_align_dir, file_ids)
with open(mlf, 'r') as f:
# Consume the MLF #!header!# line.
_ = f.readline()
for file_id in file_ids:
# Consume the file name line.
line = f.readline()
mlf_base_name = os.path.splitext(os.path.basename(line))[0]
id_base_name = os.path.basename(file_id)
if mlf_base_name != id_base_name:
raise ValueError(
f'The file order in the mlf ({mlf}) does not match file_ids)\n'
f'{mlf_base_name} {id_base_name}')
label_no_align = file_io.load_lines(
os.path.join(lab_dir, f'{file_id}.lab'))
label_state_align = []
for label_tag in label_no_align:
label_tag = label_tag.strip()
for i in range(STATES_PER_PHONE):
# Consume a state alignment line.
line = f.readline().strip()
# Get the alignments for this state.
start_time, end_time, *_ = line.split()
label_state_align.append(
f'{start_time} {end_time} {label_tag}[{i + 2}]')
# label_state_align
file_io.save_lines(
label_state_align,
os.path.join(lab_align_dir, f'{file_id}.lab'))
# Consume the end of file line marker ('.' character).
line = f.readline().strip()
if line != '.':
raise ValueError('The two files are not matched!')
def sanitise_labs(lab_dir,
file_ids,
label_out_dir,
include_times=False,
state_level=False,
is_mono=False):
utils.make_dirs(label_out_dir, file_ids)
for file_id in file_ids:
label = file_io.load_lines(os.path.join(lab_dir, f'{file_id}.lab'))
n_phones = len(label)
start_times, end_times, label = map(list, zip(*map(str.split, label)))
start_times, end_times, label = sanitise_silences(start_times,
end_times,
label,
is_mono=is_mono)
if state_level:
if include_times:
n_states = n_phones * STATES_PER_PHONE
times = np.interp(range(0, n_states + 1, 1),
range(0, n_states + 1, STATES_PER_PHONE),
start_times + end_times[-1:])
start_times = times[:-1]
end_times = times[1:]
label = np.repeat(label, STATES_PER_PHONE).tolist()
for i in range(len(label)):
state_idx = i % STATES_PER_PHONE
label[i] += f'[{state_idx+2}]'
if include_times:
start_times = list(map(_round_dur, start_times))
end_times = list(map(_round_dur, end_times))
label = list(map(' '.join, zip(*[start_times, end_times, label])))
file_io.save_lines(label, os.path.join(label_out_dir,
f'{file_id}.lab'))
def batch_synth(lf0,
vuv,
mcep,
bap,
seq_len=None,
names=None,
out_dir=None,
sample_rate=16000):
if out_dir is not None:
if names is None:
raise ValueError(
'If `out_dir` is given, then `names` of individual sentences must also be given'
)
synth_dir = os.path.join(out_dir, 'synth')
make_dirs(synth_dir, names)
lf0, vuv, mcep, bap = utils.detach_batched_seqs(lf0,
vuv,
mcep,
bap,
seq_len=seq_len)
wavs = []
for i, name in enumerate(names):
f0_i = np.exp(lf0[i])
f0_i = savgol_filter(f0_i, 7, 1) if len(f0_i) >= 7 else f0_i
wav = world.synthesis(f0_i,
vuv[i],
mcep[i],
bap[i],
sample_rate=sample_rate)
wavs.append(wav)
if out_dir is not None:
wav_path = os.path.join(synth_dir, f'{names[i]}.wav')
file_io.save_wav(wav, wav_path, sample_rate=sample_rate)
return wavs
def dumps_to_labs(dump_dir, file_ids, label_out_dir, awk='label-full.awk'):
if awk in pkg_resources.resource_listdir(
'tts_data_tools', os.path.join('resources', 'festival')):
print(
f'Using tts_data_tools resource from resources/festival for {awk}')
awk = pkg_resources.resource_filename(
'tts_data_tools', os.path.join('resources', 'festival', awk))
utils.make_dirs(label_out_dir, file_ids)
for file_id in file_ids:
# Argument `check=True` ensures that an exception is raised if the process' return code is non-zero.
rtn = subprocess.run(
['gawk', '-f', awk,
os.path.join(dump_dir, f'{file_id}.txt')],
check=True,
stdout=subprocess.PIPE)
# `stdout` was redirected with a pipe and stored in the return object `rtn` as a binary string.
with open(os.path.join(label_out_dir, f'{file_id}.lab'), 'wb') as f:
f.write(rtn.stdout)
def save_files(self, data, base_names, data_dir):
utils.make_dirs(data_dir, base_names)
for datum, base_name in zip(data, base_names):
self.save_file(datum, base_name, data_dir)
def process(lab_dir, id_list, out_dir, state_level, lab_dir_with_pos, wav_dir):
"""Processes label files in id_list, saves the phone identities (as a string) to text files.
Args:
lab_dir (str): Directory containing the label files.
id_list (str): List of file basenames to process.
out_dir (str): Directory to save the output to.
state_level (bool): Indicates that the label files are state level if True, otherwise they are frame level.
"""
file_ids = utils.get_file_ids(id_list=id_list)
utils.make_dirs(os.path.join(out_dir, 'segment_n_phones'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'segment_n_frames'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'n_segments'), file_ids)
for file_id in file_ids:
lab_path_with_pos = os.path.join(lab_dir_with_pos, f'{file_id}.lab')
label_with_pos = file_io.load_lines(lab_path_with_pos)
word_start_idxs, _ = get_word_idxs(
label_with_pos, word_idx_sep=(r'@', r'\+'), phrase_idx_sep=(r'@', r'='))
pos_tags = get_pos_tags(label_with_pos, word_start_idxs)
lab_path = os.path.join(lab_dir, f'{file_id}.lab')
label = lab_to_feat.Label(lab_path, state_level)
durations = label.phone_durations
n_frames = np.sum(durations).item()
n_phones = len(label.phones)
word_start_idxs, word_end_idxs = get_word_idxs(
label.labels, word_idx_sep=(r':', r'\+'), phrase_idx_sep=(r':', r'='))
try:
segment_start_idxs, segment_end_idxs = segment_words(word_start_idxs, word_end_idxs, pos_tags)
except (ValueError, IndexError) as e:
print(f'{e}\n{file_id}')
else:
wav_path = os.path.join(wav_dir, f'{file_id}.wav')
wav, sample_rate = file_io.load_wav(wav_path)
f0, _, _, _ = world_with_reaper_f0.analysis(wav, sample_rate)
# Match the number of frames between label forced-alignment and vocoder analysis.
# Often the durations from forced alignment are a few frames longer than the vocoder features.
diff = n_frames - f0.shape[0]
if diff > n_phones:
raise ValueError(f'Number of label frames and vocoder frames is too different for {file_id}\n'
f'\tlabel frames {n_frames}\n'
f'\tvocoder frames {f0.shape[0]}\n'
f'\tnumber of phones {n_phones}')
# Remove excess durations if there is a shape mismatch.
if diff > 0:
# Remove 1 frame from each phone's duration starting at the end of the sequence.
durations[-diff:] -= 1
n_frames = f0.shape[0]
print(f'Cropped {diff} frames from durations for utterance {file_id}')
assert n_frames == np.sum(durations).item()
segment_phone_lens = []
segment_frame_lens = []
for segment_start_idx, segment_end_idx in zip(segment_start_idxs, segment_end_idxs):
segment_phone_lens.append(segment_end_idx - segment_start_idx)
segment_frame_lens.append(sum(durations[segment_start_idx:segment_end_idx]))
file_io.save_txt(segment_phone_lens, os.path.join(out_dir, 'segment_n_phones', f'{file_id}.txt'))
file_io.save_txt(segment_frame_lens, os.path.join(out_dir, 'segment_n_frames', f'{file_id}.txt'))
file_io.save_txt(len(segment_phone_lens), os.path.join(out_dir, 'n_segments', f'{file_id}.txt'))
def utts_to_dumps(dumpfeats_exe,
utt_dir,
file_ids,
dump_dir,
feature_level='Segment',
extra_feats_scm='extra_feats.scm',
label_feats='label.feats',
custom_voice=None):
if extra_feats_scm in pkg_resources.resource_listdir(
'tts_data_tools', os.path.join('resources', 'festival')):
print(
f'Using tts_data_tools resource from resources/festival for {extra_feats_scm}'
)
extra_feats_scm = pkg_resources.resource_filename(
'tts_data_tools',
os.path.join('resources', 'festival', extra_feats_scm))
if custom_voice is not None:
# Create a temporary file, to which we can add a command to load the custom voice.
extra_feats_scm_with_custom_voice = tempfile.NamedTemporaryFile(
suffix='.scm')
# Write an initial line to load the custom voice.
extra_feats_scm_with_custom_voice.write(f'(voice_{custom_voice})\n')
# Write the code from the original Scheme file.
with open(extra_feats_scm, 'r') as f:
scm_code = f.read()
extra_feats_scm_with_custom_voice.write(scm_code)
# Replace the file name with the name of the temporary file.
extra_feats_scm = extra_feats_scm_with_custom_voice.name
if label_feats in pkg_resources.resource_listdir(
'tts_data_tools', os.path.join('resources', 'festival')):
print(
f'Using tts_data_tools resource from resources/festival for {label_feats}'
)
label_feats = pkg_resources.resource_filename(
'tts_data_tools', os.path.join('resources', 'festival',
label_feats))
utils.make_dirs(dump_dir, file_ids)
for file_id in file_ids:
# Argument `check=True` ensures that an exception is raised if the process' return code is non-zero.
subprocess.run([
dumpfeats_exe, '-eval', extra_feats_scm, '-relation',
feature_level, '-feats', label_feats, '-output',
os.path.join(dump_dir, f'{file_id}.txt'),
os.path.join(utt_dir, f'{file_id}.utt')
],
check=True)
# Replace any '#' characters used for pauses with 'pau'.
subprocess.run([
'sed', '-i', '-e', 's/#/pau/g',
*glob.glob('label_POS/label_phone_align/dump/*')
],
check=True)
if custom_voice is not None:
# Make sure to close the temporary file, this ensures it gets deleted.
extra_feats_scm_with_custom_voice.close()
def process(lab_dir, wav_dir, id_list, out_dir, state_level, question_file,
upsample, subphone_feat_type, trim_silences,
calculate_normalisation, normalisation_of_deltas):
"""Processes wav files in id_list, saves the log-F0 and MVN parameters to files.
Args:
lab_dir (str): Directory containing the label files.
wav_dir (str): Directory containing the wav files.
id_list (str): List of file basenames to process.
out_dir (str): Directory to save the output to.
state_level (bool): Indicates that the label files are state level if True, otherwise they are frame level.
question_file (str): Question set to be loaded. Can be one of the four provided question sets;
questions-unilex_dnn_600.hed
questions-unilex_phones_69.hed
questions-radio_dnn_416.hed
questions-radio_phones_48.hed
questions-mandarin.hed
questions-japanese.hed
upsample (bool): Whether to upsample phone-level numerical labels to frame-level.
subphone_feat_type (str): Subphone features to be extracted from the durations.
trim_silences (bool): Whether to trim start and end silences from all features.
calculate_normalisation (bool): Whether to automatically calculate MVN parameters after extracting F0.
normalisation_of_deltas (bool): Also calculate the MVN parameters for the delta and delta delta features.
"""
file_ids = utils.get_file_ids(id_list=id_list)
question_set = lab_to_feat.QuestionSet(question_file)
subphone_feature_set = lab_to_feat.SubphoneFeatureSet(subphone_feat_type)
utils.make_dirs(os.path.join(out_dir, 'lab'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'counters'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'dur'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'phones'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'n_frames'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'n_phones'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'lf0'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'vuv'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'mcep'), file_ids)
utils.make_dirs(os.path.join(out_dir, 'bap'), file_ids)
for file_id in tqdm(file_ids):
# Label processing.
lab_path = os.path.join(lab_dir, f'{file_id}.lab')
label = lab_to_feat.Label(lab_path, state_level)
numerical_labels = label.extract_numerical_labels(
question_set, upsample_to_frame_level=upsample)
counter_features = label.extract_counter_features(subphone_feature_set)
durations = label.phone_durations.reshape((-1, 1))
phones = label.phones
n_frames = np.sum(durations).item()
n_phones = len(label.phones)
# Acoustic processing.
wav_path = os.path.join(wav_dir, f'{file_id}.wav')
wav, sample_rate = file_io.load_wav(wav_path)
f0, vuv, mcep, bap = world_with_reaper_f0.analysis(wav, sample_rate)
lf0 = np.log(f0)
# Match the number of frames between label forced-alignment and vocoder analysis.
# Often the durations from forced alignment are a few frames longer than the vocoder features.
diff = n_frames - f0.shape[0]
if diff > n_phones:
raise ValueError(
f'Number of label frames and vocoder frames is too different for {file_id}\n'
f'\tlabel frames {n_frames}\n'
f'\tvocoder frames {f0.shape[0]}\n'
f'\tnumber of phones {n_phones}')
# Remove excess durations if there is a shape mismatch.
if diff > 0:
# Remove 1 frame from each phone's duration starting at the end of the sequence.
durations[-diff:] -= 1
n_frames = f0.shape[0]
print(
f'Cropped {diff} frames from durations for utterance {file_id}'
)
assert n_frames == np.sum(durations).item()
trim_frame_slice = slice(0, n_frames)
if trim_silences:
start_phone_idx, end_phone_idx = 0, n_phones
start_frame_idx, end_frame_idx = 0, n_frames
if phones[0] in ['sil', '#']:
start_phone_idx += 1
start_frame_idx += durations[0]
if phones[-1] in ['sil', '#']:
end_phone_idx -= 1
end_frame_idx -= durations[-1]
trim_phone_slice = slice(int(start_phone_idx), int(end_phone_idx))
trim_frame_slice = slice(int(start_frame_idx), int(end_frame_idx))
numerical_labels = numerical_labels[
trim_frame_slice if upsample else trim_phone_slice]
durations = durations[trim_phone_slice]
phones = phones[trim_phone_slice]
n_frames = trim_frame_slice.stop - trim_frame_slice.start
n_phones = trim_phone_slice.stop - trim_phone_slice.start
counter_features = counter_features[trim_frame_slice]
lf0 = lf0[trim_frame_slice]
vuv = vuv[trim_frame_slice]
mcep = mcep[trim_frame_slice]
bap = bap[trim_frame_slice]
file_io.save_bin(numerical_labels.astype(np.float32),
os.path.join(out_dir, 'lab', f'{file_id}.npy'))
file_io.save_bin(counter_features.astype(np.float32),
os.path.join(out_dir, 'counters', f'{file_id}.npy'))
file_io.save_txt(durations,
os.path.join(out_dir, 'dur', f'{file_id}.txt'))
file_io.save_lines(phones,
os.path.join(out_dir, 'phones', f'{file_id}.txt'))
file_io.save_txt(n_frames,
os.path.join(out_dir, 'n_frames', f'{file_id}.txt'))
file_io.save_txt(n_phones,
os.path.join(out_dir, 'n_phones', f'{file_id}.txt'))
file_io.save_bin(lf0.astype(np.float32),
os.path.join(out_dir, 'lf0', f'{file_id}.npy'))
file_io.save_bin(vuv, os.path.join(out_dir, 'vuv', f'{file_id}.npy'))
file_io.save_bin(mcep.astype(np.float32),
os.path.join(out_dir, 'mcep', f'{file_id}.npy'))
file_io.save_bin(bap.astype(np.float32),
os.path.join(out_dir, 'bap', f'{file_id}.npy'))
if calculate_normalisation:
process_minmax(out_dir, 'lab', id_list, out_dir=out_dir)
process_minmax(out_dir, 'counters', id_list, out_dir=out_dir)
process_mvn(out_dir,
'dur',
is_npy=False,
id_list=id_list,
deltas=False,
out_dir=out_dir)
process_mvn(out_dir,
'lf0',
id_list=id_list,
deltas=normalisation_of_deltas,
out_dir=out_dir)
process_mvn(out_dir,
'mcep',
id_list=id_list,
deltas=normalisation_of_deltas,
out_dir=out_dir)
process_mvn(out_dir,
'bap',
id_list=id_list,
deltas=normalisation_of_deltas,
out_dir=out_dir)
def process(lab_dir, id_list, out_dir, state_level, question_file, upsample,
subphone_feat_type, calculate_normalisation):
"""Processes label files in id_list, saves the numerical labels and durations to file.
Args:
lab_dir (str): Directory containing the label files.
id_list (str): List of file basenames to process.
out_dir (str): Directory to save the output to.
state_level (bool): Indicates that the label files are state level if True, otherwise they are frame level.
question_file (str): Question set to be loaded. Can be one of the four provided question sets;
questions-unilex_dnn_600.hed
questions-unilex_phones_69.hed
questions-radio_dnn_416.hed
questions-radio_phones_48.hed
questions-mandarin.hed
questions-japanese.hed
upsample (bool): Whether to upsample phone-level numerical labels to frame-level.
subphone_feat_type (str): Subphone features to be extracted from the durations.
calculate_normalisation (bool): Calculate mean-variance and min-max normalisation for duration and labels.
"""
file_ids = get_file_ids(id_list=id_list)
question_set = QuestionSet(question_file)
subphone_feature_set = SubphoneFeatureSet(subphone_feat_type)
make_dirs(os.path.join(out_dir, 'lab'), file_ids)
make_dirs(os.path.join(out_dir, 'counters'), file_ids)
make_dirs(os.path.join(out_dir, 'dur'), file_ids)
make_dirs(os.path.join(out_dir, 'phones'), file_ids)
make_dirs(os.path.join(out_dir, 'n_frames'), file_ids)
make_dirs(os.path.join(out_dir, 'n_phones'), file_ids)
for file_id in file_ids:
lab_path = os.path.join(lab_dir, f'{file_id}.lab')
label = Label(lab_path, state_level)
numerical_labels = label.extract_numerical_labels(
question_set, upsample_to_frame_level=upsample)
counter_features = label.extract_counter_features(subphone_feature_set)
durations = label.phone_durations.reshape((-1, 1))
phones = label.phones
n_frames = np.sum(durations).item()
n_phones = len(label.phones)
file_io.save_bin(numerical_labels,
os.path.join(out_dir, 'lab', f'{file_id}.npy'))
file_io.save_bin(counter_features,
os.path.join(out_dir, 'counters', f'{file_id}.npy'))
file_io.save_txt(durations,
os.path.join(out_dir, 'dur', f'{file_id}.dur'))
file_io.save_lines(phones,
os.path.join(out_dir, 'phones', f'{file_id}.txt'))
file_io.save_txt(n_frames,
os.path.join(out_dir, 'n_frames', f'{file_id}.txt'))
file_io.save_txt(n_phones,
os.path.join(out_dir, 'n_phones', f'{file_id}.txt'))
if calculate_normalisation:
process_minmax(out_dir, 'lab', id_list)
process_minmax(out_dir, 'counters', id_list)
process_mvn(out_dir,
'dur',
is_npy=False,
id_list=id_list,
deltas=False)