Update information

[kan-bayashi]

2018/05/01

• Updated to be compatible with pytorch v0.4
• Updated to be able to use melspectrogram as auxiliary feature

Due to above update, some parts are changed (see below)

# -------------------- #
# feature path in hdf5 #
# -------------------- #
old -> new
/feat_org -> /world or /melspc
/feat -> no more saving extended featrue (it is replicated when loading)

# ----------------------- #
# statistics path in hdf5 #
# ----------------------- #
old -> new
/mean -> /world/mean or /melspc/mean
/scale -> /world/scale or /melspc/scale

# ----------------------- #
# new options in training #
# ----------------------- #
--feature_type: Auxiliary feature type (world or melspc)
--use_upsampling_layer: Flag to decide whether to use upsampling layer in WaveNet
--upsampling_factor: Changed to be alway needed because feature extension is performed in loading

Note that old model file checkpoint-*.pkl can be used, but it is necessary to modify model.conf file as follows.

# how-to-convert to new config file
import torch
args = torch.load("old_model.conf")
args.use_upsampling_layer = True
args.feature_type = "world"
torch.save(args, "new_model.conf")

[kan-bayashi]

2018/05/07

• Added samples using mel-spectrogram

[kan-bayashi]

2018/05/11

• Implemented noise-shaping stft-based mel-cepstrum
• All of the recipes can use noise shaping technique
• Added sample of mel-spectrogram with noise shaping

[kan-bayashi]

2018/06/21

• Chaned mel-spectrogram calculation to use amp spec instead of power spec
• Tacotron2 + WaveNet vocoder (without joint-training) samples are ready! https://drive.google.com/drive/folders/1uwuakmn1fyfO7IQ8Sq51yfq89CL1UHlh?usp=sharing