def __init__(self, params, wavenet):
self.params = params
self.UPSAMPLE_COUNT = int(12.5 * params.target_sample_rate / 1000)
self.RECEPTIVE_SIZE = 512 # this means 16ms
self.MGC_SIZE = params.mgc_order
self.dio = DatasetIO()
self.vocoder = MelVocoder()
self.wavenet = wavenet
self.network = ParallelVocoderNetwork(receptive_field=self.RECEPTIVE_SIZE, filter_size=64).to(device)
self.trainer = torch.optim.Adam(self.network.parameters(), lr=self.params.learning_rate)
def __init__(self, params, model=None, runtime=False):
self.params = params
self.HIDDEN_SIZE = [1000, 1000]
self.FFT_SIZE = 513
self.UPSAMPLE_COUNT = int(12.5 * params.target_sample_rate / 1000)
self.FILTER_SIZE = 128
self.sparse = False
if model is None:
self.model = dy.Model()
else:
self.model = model
input_size = self.params.mgc_order # self.UPSAMPLE_COUNT + self.params.mgc_order
hidden_w = []
hidden_b = []
for layer_size in self.HIDDEN_SIZE:
hidden_w.append(self.model.add_parameters((layer_size, input_size)))
hidden_b.append(self.model.add_parameters((layer_size)))
input_size = layer_size
self.mlp_excitation = [hidden_w, hidden_b]
input_size = self.params.mgc_order # self.UPSAMPLE_COUNT + self.params.mgc_order
hidden_w = []
hidden_b = []
for layer_size in self.HIDDEN_SIZE:
hidden_w.append(self.model.add_parameters((layer_size, input_size)))
hidden_b.append(self.model.add_parameters((layer_size)))
input_size = layer_size
self.mlp_filter = [hidden_w, hidden_b]
input_size = self.params.mgc_order # self.UPSAMPLE_COUNT + self.params.mgc_order
hidden_w = []
hidden_b = []
for layer_size in self.HIDDEN_SIZE:
hidden_w.append(self.model.add_parameters((layer_size, input_size)))
hidden_b.append(self.model.add_parameters((layer_size)))
input_size = layer_size
self.mlp_vuv = [hidden_w, hidden_b]
self.excitation_w = self.model.add_parameters((self.UPSAMPLE_COUNT + self.FILTER_SIZE - 1, input_size))
self.excitation_b = self.model.add_parameters((self.UPSAMPLE_COUNT + self.FILTER_SIZE - 1))
self.filter_w = self.model.add_parameters((self.FILTER_SIZE, input_size))
self.filter_b = self.model.add_parameters((self.FILTER_SIZE))
self.vuv_w = self.model.add_parameters((1, input_size))
self.vuv_b = self.model.add_parameters((1))
self.trainer = dy.AdamTrainer(self.model, alpha=params.learning_rate)
self.dio = DatasetIO()
self.vocoder = MelVocoder()
def __init__(self, params, model=None, runtime=False):
self.params = params
self.UPSAMPLE_COUNT = int(12.5 * params.target_sample_rate / 1000)
self.RECEPTIVE_SIZE = 512 # this means 16ms
self.dio = DatasetIO()
self.vocoder = MelVocoder()
self.network = VocoderNetwork(receptive_field=self.RECEPTIVE_SIZE, filter_size=256).to(device)
self.trainer = torch.optim.Adam(self.network.parameters(), lr=self.params.learning_rate)
def __init__(self, params, model=None, runtime=False):
self.params = params
self.HIDDEN_SIZE = [1000, 1000]
self.FFT_SIZE = 513
self.UPSAMPLE_COUNT = int(12.5 * params.target_sample_rate / 1000)
self.FILTER_SIZE = 128
self.sparse = False
self.dio = DatasetIO()
self.vocoder = MelVocoder()
self.network = VocoderNetwork(self.params.mgc_order,
self.UPSAMPLE_COUNT).to(device)
self.trainer = torch.optim.Adam(self.network.parameters(),
lr=self.params.learning_rate)
def phase_1_prepare_corpus(params):
from os import listdir
from os.path import isfile, join
from os.path import exists
train_files_tmp = [
f for f in listdir(params.train_folder)
if isfile(join(params.train_folder, f))
]
dev_files_tmp = [
f for f in listdir(params.dev_folder)
if isfile(join(params.dev_folder, f))
]
sys.stdout.write("Scanning training files...")
sys.stdout.flush()
final_list = []
for file in train_files_tmp:
base_name = file[:-4]
lab_name = base_name + '.txt'
wav_name = base_name + '.wav'
if exists(join(params.train_folder, lab_name)) and exists(
join(params.train_folder, wav_name)):
if base_name not in final_list:
final_list.append(base_name)
train_files = final_list
sys.stdout.write(" found " + str(len(train_files)) +
" valid training files\n")
sys.stdout.write("Scanning development files...")
sys.stdout.flush()
final_list = []
for file in dev_files_tmp:
base_name = file[:-4]
lab_name = base_name + '.txt'
wav_name = base_name + '.wav'
if exists(join(params.dev_folder, lab_name)) and exists(
join(params.dev_folder, wav_name)):
if base_name not in final_list:
final_list.append(base_name)
dev_files = final_list
sys.stdout.write(" found " + str(len(dev_files)) +
" valid development files\n")
from io_modules.dataset import DatasetIO
from io_modules.vocoder import MelVocoder
from shutil import copyfile
import pysptk
dio = DatasetIO()
vocoder = MelVocoder()
base_folder = params.train_folder
for index in range(len(train_files)):
sys.stdout.write("\r\tprocessing file " + str(index + 1) + "/" +
str(len(train_files)))
sys.stdout.flush()
base_name = train_files[index]
txt_name = base_name + '.txt'
wav_name = base_name + '.wav'
spc_name = base_name + '.png'
lab_name = base_name + '.lab'
# LAB - copy or create
if exists(join(base_folder, lab_name)):
copyfile(join(base_folder, lab_name),
join('data/processed/train', lab_name))
else:
create_lab_file(join(base_folder, txt_name),
join('data/processed/train', lab_name))
# TXT
copyfile(join(base_folder, txt_name),
join('data/processed/train', txt_name))
# WAVE
data, sample_rate = dio.read_wave(
join(base_folder, wav_name),
sample_rate=params.target_sample_rate)
mgc = vocoder.melspectrogram(data,
sample_rate=params.target_sample_rate,
num_mels=params.mgc_order)
# SPECT
render_spectrogram(mgc, join('data/processed/train', spc_name))
dio.write_wave(
join('data/processed/train', base_name + '.orig.wav'), data,
sample_rate)
array2file(mgc, join('data/processed/train', base_name + '.mgc'))
sys.stdout.write('\n')
base_folder = params.dev_folder
for index in range(len(dev_files)):
sys.stdout.write("\r\tprocessing file " + str(index + 1) + "/" +
str(len(dev_files)))
sys.stdout.flush()
base_name = dev_files[index]
txt_name = base_name + '.txt'
wav_name = base_name + '.wav'
spc_name = base_name + '.png'
lab_name = base_name + '.lab'
# LAB - copy or create
if exists(join(base_folder, lab_name)):
copyfile(join(base_folder, lab_name),
join('data/processed/dev', lab_name))
else:
create_lab_file(join(base_folder, txt_name),
join('data/processed/dev', lab_name))
# TXT
copyfile(join(base_folder, txt_name),
join('data/processed/dev/', txt_name))
# WAVE
data, sample_rate = dio.read_wave(
join(base_folder, wav_name),
sample_rate=params.target_sample_rate)
mgc = vocoder.melspectrogram(data,
sample_rate=params.target_sample_rate,
num_mels=params.mgc_order)
# SPECT
render_spectrogram(mgc, join('data/processed/dev', spc_name))
dio.write_wave(join('data/processed/dev', base_name + '.orig.wav'),
data, sample_rate)
array2file(mgc, join('data/processed/dev', base_name + '.mgc'))
sys.stdout.write('\n')
def phase_1_prepare_corpus(params):
from os import listdir
from os.path import isfile, join
from os.path import exists
train_files_tmp = [
f for f in listdir(params.train_folder)
if isfile(join(params.train_folder, f))
]
if params.dev_folder is not None:
dev_files_tmp = [
f for f in listdir(params.dev_folder)
if isfile(join(params.dev_folder, f))
]
else:
dev_files_tmp = []
if params.g2p is not None:
from models.g2p import G2P
from io_modules.encodings import Encodings
g2p_encodings = Encodings()
g2p_encodings.load(params.g2p + '.encodings')
g2p = G2P(g2p_encodings)
g2p.load(params.g2p + '-bestAcc.network')
if exists(params.g2p + '.lexicon'):
g2p.load_lexicon(params.g2p + '.lexicon')
else:
g2p = None
sys.stdout.write("Scanning training files...")
sys.stdout.flush()
final_list = []
for file in train_files_tmp:
base_name = file[:-4]
lab_name = base_name + '.txt'
wav_name = base_name + '.wav'
if exists(join(params.train_folder, lab_name)) and exists(
join(params.train_folder, wav_name)):
if base_name not in final_list:
final_list.append(base_name)
train_files = final_list
sys.stdout.write(" found " + str(len(train_files)) +
" valid training files\n")
sys.stdout.write("Scanning development files...")
sys.stdout.flush()
final_list = []
for file in dev_files_tmp:
base_name = file[:-4]
lab_name = base_name + '.txt'
wav_name = base_name + '.wav'
if exists(join(params.dev_folder, lab_name)) and exists(
join(params.dev_folder, wav_name)):
if base_name not in final_list:
final_list.append(base_name)
dev_files = final_list
sys.stdout.write(" found " + str(len(dev_files)) +
" valid development files\n")
from io_modules.dataset import DatasetIO
from io_modules.vocoder import MelVocoder
from shutil import copyfile
dio = DatasetIO()
vocoder = MelVocoder()
base_folder = params.train_folder
total_files = 0
for index in range(len(train_files)):
total_files += 1
sys.stdout.write("\r\tprocessing file " + str(index + 1) + "/" +
str(len(train_files)))
sys.stdout.flush()
base_name = train_files[index]
txt_name = base_name + '.txt'
wav_name = base_name + '.wav'
spc_name = base_name + '.png'
lab_name = base_name + '.lab'
tgt_txt_name = txt_name
tgt_spc_name = spc_name
tgt_lab_name = lab_name
if params.prefix is not None:
tgt_txt_name = params.prefix + "_{:05d}".format(
total_files) + '.txt'
tgt_spc_name = params.prefix + "_{:05d}".format(
total_files) + '.png'
tgt_lab_name = params.prefix + "_{:05d}".format(
total_files) + '.lab'
# LAB - copy or create
if exists(join(base_folder, lab_name)):
copyfile(join(base_folder, lab_name),
join('data/processed/train', tgt_lab_name))
else:
create_lab_file(join(base_folder, txt_name),
join('data/processed/train', tgt_lab_name),
speaker_name=params.speaker,
g2p=g2p)
# TXT
copyfile(join(base_folder, txt_name),
join('data/processed/train', tgt_txt_name))
# WAVE
data, sample_rate = dio.read_wave(
join(base_folder, wav_name),
sample_rate=params.target_sample_rate)
mgc = vocoder.melspectrogram(data,
sample_rate=params.target_sample_rate,
num_mels=params.mgc_order)
# SPECT
render_spectrogram(mgc, join('data/processed/train', tgt_spc_name))
if params.prefix is None:
dio.write_wave(
join('data/processed/train', base_name + '.orig.wav'),
data, sample_rate)
array2file(mgc, join('data/processed/train',
base_name + '.mgc'))
else:
tgt_wav_name = params.prefix + "_{:05d}".format(
total_files) + '.orig.wav'
tgt_mgc_name = params.prefix + "_{:05d}".format(
total_files) + '.mgc'
dio.write_wave(join('data/processed/train', tgt_wav_name),
data, sample_rate)
array2file(mgc, join('data/processed/train', tgt_mgc_name))
sys.stdout.write('\n')
base_folder = params.dev_folder
for index in range(len(dev_files)):
total_files += 1
sys.stdout.write("\r\tprocessing file " + str(index + 1) + "/" +
str(len(dev_files)))
sys.stdout.flush()
base_name = dev_files[index]
txt_name = base_name + '.txt'
wav_name = base_name + '.wav'
spc_name = base_name + '.png'
lab_name = base_name + '.lab'
tgt_txt_name = txt_name
tgt_spc_name = spc_name
tgt_lab_name = lab_name
if params.prefix is not None:
tgt_txt_name = params.prefix + "_{:05d}".format(
total_files) + '.txt'
tgt_spc_name = params.prefix + "_{:05d}".format(
total_files) + '.png'
tgt_lab_name = params.prefix + "_{:05d}".format(
total_files) + '.lab'
# LAB - copy or create
if exists(join(base_folder, lab_name)):
copyfile(join(base_folder, lab_name),
join('data/processed/dev', tgt_lab_name))
else:
create_lab_file(join(base_folder, txt_name),
join('data/processed/dev', tgt_lab_name),
speaker_name=params.speaker,
g2p=g2p)
# TXT
copyfile(join(base_folder, txt_name),
join('data/processed/dev', tgt_txt_name))
# WAVE
data, sample_rate = dio.read_wave(
join(base_folder, wav_name),
sample_rate=params.target_sample_rate)
mgc = vocoder.melspectrogram(data,
sample_rate=params.target_sample_rate,
num_mels=params.mgc_order)
# SPECT
render_spectrogram(mgc, join('data/processed/dev', tgt_spc_name))
if params.prefix is None:
dio.write_wave(
join('data/processed/dev', base_name + '.orig.wav'), data,
sample_rate)
array2file(mgc, join('data/processed/dev', base_name + '.mgc'))
else:
tgt_wav_name = params.prefix + "_{:05d}".format(
total_files) + '.orig.wav'
tgt_mgc_name = params.prefix + "_{:05d}".format(
total_files) + '.mgc'
dio.write_wave(join('data/processed/dev', tgt_wav_name), data,
sample_rate)
array2file(mgc, join('data/processed/dev', tgt_mgc_name))
sys.stdout.write('\n')