def __init__(self):
speakers = list()
for file in os.listdir(os.path.join(hparams.VCTK_dataset_path,
"wav48")):
speakers.append(str(file[1:4]))
self.speaker_list = speakers
# Shuffle Speakers
if not os.path.exists(hparams.shuffle_speaker_file):
shuffle_list = [i for i in range(len(self.speaker_list))]
random.shuffle(shuffle_list)
shuffle_list = shuffle_list[0:hparams.shuffle_speaker_num]
self.speaker_list = self.shuffle(self.speaker_list, shuffle_list)
with open(hparams.shuffle_speaker_file, "w") as f:
for speaker in shuffle_list:
f.write(str(speaker) + "\n")
else:
shuffle_list = list()
with open(hparams.shuffle_speaker_file, "r") as f:
for ele in f.readlines():
shuffle_list.append(int(ele))
self.speaker_list = self.shuffle(self.speaker_list, shuffle_list)
td = vctk.TranscriptionDataSource(hparams.VCTK_dataset_path,
speakers=self.speaker_list)
transcriptions = td.collect_files()
wav_paths = vctk.WavFileDataSource(
hparams.VCTK_dataset_path,
speakers=self.speaker_list).collect_files()
self.text = transcriptions
self.wav_paths = wav_paths
def build_from_path(in_dir, out_dir, num_workers=1, tqdm=lambda x: x):
executor = ProcessPoolExecutor(max_workers=num_workers)
futures = []
speakers = vctk.available_speakers
td = vctk.TranscriptionDataSource(in_dir, speakers=speakers)
transcriptions = td.collect_files()
speaker_ids = td.labels
speaker_ids_unique = np.unique(speaker_ids)
speaker_to_speaker_id = {}
for i, j in zip(speakers, speaker_ids_unique):
speaker_to_speaker_id[i] = j
wav_paths = vctk.WavFileDataSource(in_dir,
speakers=speakers).collect_files()
_ignore_speaker = hparams.not_for_train_speaker.split(", ")
ignore_speaker = [speaker_to_speaker_id[i] for i in _ignore_speaker]
for index, (speaker_id, text, wav_path) in enumerate(
zip(speaker_ids, transcriptions, wav_paths)):
if speaker_id in ignore_speaker:
continue
futures.append(
executor.submit(
partial(_process_utterance, out_dir, index + 1, speaker_id,
wav_path, text)))
return [future.result() for future in tqdm(futures)]
def __init__(self, data_root='/home/zeng/work/data/VCTK-Corpus/'):
super(VCTKDatasets, self).__init__()
data_sources = vctk.WavFileDataSource(data_root)
self.paths = data_sources.collect_files()
data_sources.speakers.sort()
self.speakers = data_sources.speakers[:10]
self.train_speaker = self.speakers[:int(len(self.speakers) * 0.9)]
self.test_speaker = self.speakers[int(len(self.speakers) * 0.9):]
self.datasets = {_id: SpeakerDatasets(_id, list(filter(lambda x: 'p' + _id in x, self.paths))) for _id in
self.speakers}
def preprocess():
speakers = vctk.available_speakers
wav_paths = vctk.WavFileDataSource(
hp.origin_data, speakers=speakers).collect_files()
if not os.path.exists("dataset"):
os.mkdir("dataset")
cnt_speaker = -1
cnt_num = -1
dict_speaker = list()
num_dict = list()
mel_spec_list = list()
for wav_file in wav_paths:
base_name = os.path.basename(wav_file)
speaker_id = int(base_name[1:4])
cnt_id = int(base_name[5:8])
if speaker_id not in dict_speaker:
dict_speaker.append(speaker_id)
cnt_speaker = cnt_speaker + 1
num_dict.clear()
cnt_num = -1
if cnt_id not in num_dict:
num_dict.append(cnt_id)
cnt_num = cnt_num + 1
spec_name = str(cnt_speaker) + "_" + str(cnt_num) + ".npy"
mel_spec_list.append(spec_name)
# executor = ProcessPoolExecutor(max_workers=cpu_count())
# futures = list()
wav_temp_list = list()
save_temp_list = list()
total_len = len(wav_paths)
for ind, wav_file in enumerate(wav_paths):
wav_temp_list.append(wav_file)
save_temp_list.append(mel_spec_list[ind])
if (((ind + 1) % 1000) == 0) or (ind == total_len - 1):
save_by_list(wav_temp_list.copy(), save_temp_list.copy())
# futures.append(executor.submit(
# partial(save_by_list, wav_temp_list.copy(), save_temp_list.copy())))
wav_temp_list.clear()
save_temp_list.clear()
print("Done", ind+1)
print("Done")
def __init__(self, dataset_path=hparams.dataset_path):
speakers = list()
for file in os.listdir(os.path.join(hparams.VCTK_dataset_path,
"wav48")):
speakers.append(str(file[1:4]))
self.speaker_list = speakers
# print(speakers)
td = vctk.TranscriptionDataSource(hparams.VCTK_dataset_path,
speakers=speakers)
transcriptions = td.collect_files()
wav_paths = vctk.WavFileDataSource(hparams.VCTK_dataset_path,
speakers=speakers).collect_files()
self.dataset_path = dataset_path
self.text_path = os.path.join(self.dataset_path, "train.txt")
self.text = transcriptions
self.wav_paths = wav_paths
def build_from_path(in_dir, out_dir, num_workers=1, tqdm=lambda x: x):
executor = ProcessPoolExecutor(max_workers=num_workers)
futures = []
speakers = vctk.available_speakers
td = vctk.TranscriptionDataSource(in_dir, speakers=speakers)
transcriptions = td.collect_files()
speaker_ids = td.labels
wav_paths = vctk.WavFileDataSource(
in_dir, speakers=speakers).collect_files()
for index, (speaker_id, text, wav_path) in enumerate(
zip(speaker_ids, transcriptions, wav_paths)):
futures.append(executor.submit(
partial(_process_utterance, out_dir, index + 1, speaker_id, wav_path, text)))
return [future.result() for future in tqdm(futures)]
def __init__(self, vctk_path='/home/zeng/work/data/VCTK-Corpus/',
data_root='data',
mode='TD-SV'):
self.data_root = data_root
self.speakers = vctk.WavFileDataSource(data_root=vctk_path)
speakers = self.speakers.labelmap.keys()
self.speakerDataSource = {'p' + speaker_id: PyTorchDataset_TISV(
_NPYDataSource(data_root=data_root,
col='',
speaker_id='p' + speaker_id,
text_index=None if mode == 'TD-SV' else '001',
train=True,
test_size=0.1
)) for speaker_id in speakers}
self.speakerDataSource_index = {}
self.speakerDataLoader = {
_id: data_utils.DataLoader(dataset=self.speakerDataSource[_id], batch_size=hparams.M,
shuffle=True)
for _id in self.speakerDataSource.keys()
}
for index, key in enumerate(self.speakerDataSource.keys()):
self.speakerDataSource_index[index] = key
from nnmnkwii.datasets import vctk
vctk_dataset = vctk.WavFileDataSource('/home/zeng/work/data/VCTK-Corpus/')
print(vctk_dataset.collect_files())
args = parser.parse_args()
log_level = args.log.upper()
logging.getLogger().setLevel(log_level)
disfluencies = set(['uh', 'um'])
data_root = args.data_root
# Do for all speakers
speakers = vctk.available_speakers
# Collect all transcripts/wav files
td = vctk.TranscriptionDataSource(data_root, speakers=speakers)
transcriptions = td.collect_files()
wav_paths = vctk.WavFileDataSource(data_root,
speakers=speakers).collect_files()
# Save dir
save_dir = join(data_root, "lab")
if not exists(save_dir):
os.makedirs(save_dir)
resources = gentle.Resources()
for idx in tqdm(range(len(wav_paths))):
transcript = transcriptions[idx]
audiofile = wav_paths[idx]
lab_path = audiofile.replace("wav48/", "lab/").replace(".wav", ".lab")
print(transcript)
print(audiofile)
print(lab_path)
def test3():
class VCTKDatasets():
def __init__(self, data_root='/home/zeng/work/data/VCTK-Corpus/'):
super(VCTKDatasets, self).__init__()
data_sources = vctk.WavFileDataSource(data_root)
self.paths = data_sources.collect_files()
data_sources.speakers.sort()
self.speakers = data_sources.speakers[:10]
self.train_speaker = self.speakers[:int(len(self.speakers) * 0.9)]
self.test_speaker = self.speakers[int(len(self.speakers) * 0.9):]
self.datasets = {_id: SpeakerDatasets(_id, list(filter(lambda x: 'p' + _id in x, self.paths))) for _id in
self.speakers}
def sample_speakers(self, N, train=True):
speakers = self.train_speaker if train else self.test_speaker
speaker_selected = random.sample(speakers, N)
return [self.datasets[_id] for _id in speaker_selected]
vctk_datasets = VCTKDatasets()
dsets = vctk_datasets.sample_speakers(3)[0]
print(dsets)
data_sources = vctk.WavFileDataSource(data_root='/home/zeng/work/data/VCTK-Corpus/')
speakers = data_sources.speakers
files = data_sources.collect_files()
print(files)
speaker_path = list(filter(lambda x: 'p295' in x, files))
sd = SpeakerDatasets('p295',speaker_path)
import os
from nnmnkwii.datasets import vctk
import hparams as hp
# speakers = vctk.available_speakers
# print(speakers)
# print(len(speakers))
speakers = list()
for file in os.listdir(os.path.join(hp.vctk_processed, "wav48")):
speakers.append(str(file[1:4]))
# print(speakers)
# print(len(speakers))
td = vctk.TranscriptionDataSource(hp.vctk_processed, speakers=speakers)
transcriptions = td.collect_files()
wav_paths = vctk.WavFileDataSource(hp.vctk_processed,
speakers=speakers).collect_files()
print(transcriptions[32306])
print(wav_paths[32306])
def prepare_txt_dict():
speakers = vctk.available_speakers
td = vctk.TranscriptionDataSource(hp.vctk_path, speakers=speakers)
transcriptions = td.collect_files()
wav_paths = vctk.WavFileDataSource(hp.vctk_path,
speakers=speakers).collect_files()
executor = ProcessPoolExecutor(max_workers=cpu_count())
futures = list()
save_name_list = list()
if not os.path.exists("processed"):
os.mkdir("processed")
for ind in range(len(wav_paths)):
savename = os.path.basename(
wav_paths[ind])[0:len(os.path.basename(wav_paths[ind])) -
4] + ".txt"
savename = os.path.join("processed", savename)
save_name_list.append(savename)
# print(savename)
print("Get Name Done.")
lists_P = list()
with g2p.Session():
for i, text in enumerate(transcriptions):
list_not_alpha = list()
for ind, ele in enumerate(text):
if (not ele.isalpha()) and (ele != ' '):
list_not_alpha.append(ind)
# print(list_not_alpha)
cnt = 0
for ind in list_not_alpha:
text = delete_alpha_str(text, ind - cnt)
cnt = cnt + 1
# print(text + "######")
# os.path.basename(wav_paths[ind])[0:len(os.path.basename(wav_paths[ind]))-4]
# print(os.path.basename(wav_paths[ind])[0:len(os.path.basename(wav_paths[ind]))-4])
list_P = g2p.g2p(text)
# print("...")
# prepare_txt(savename, text)
# futures.append(executor.submit(partial(prepare_txt, save_name_list[ind], list_P)))
lists_P.append(list_P)
if i % 100 == 0:
print(i)
print("Get P Done.")
for ind, list_P in enumerate(lists_P):
futures.append(
executor.submit(partial(prepare_txt, save_name_list[ind], list_P)))
print("Prepare Done.")
words_dict = dict()
for future in futures:
# print(future.result())
words_dict.update(future.result())
# print(word_P_dict)
with open("words_dict.txt", "w") as f:
for key in words_dict:
temp_str_P = str()
for P in words_dict[key]:
temp_str_P = temp_str_P + P + " "
str_write = key + " " + temp_str_P
f.write(str_write + "\n")