ljspeech.yaml

# data processing
p_pronunciation: 0.99
sample_rate: 22050 # Hz
n_fft: 1024
win_length: 1024
hop_length: 256
n_mels: 80
reduction_factor: 4

# model-s2s
n_speakers: 1
speaker_dim: 16
char_dim: 256
encoder_dim: 64
kernel_size: 5
encoder_layers: 7
decoder_layers: 8
prenet_sizes: [128]
attention_dim: 128

# model-postnet
postnet_layers: 5
postnet_dim: 256

# position embedding
position_weight: 1.0
position_rate: 5.54
forward_step: 4
backward_step: 0

dropout: 0.05

# output-griffinlim
sharpening_factor: 1.4

# optimizer:
learning_rate: 0.001
clip_value: 5.0
clip_norm: 100.0

# training:
max_iteration: 1000000
batch_size: 16
report_interval: 10000
save_interval: 10000
valid_size: 5