text_to_speech/egs/egs_bases/tts/ps.yaml

base_config: ./fs.yaml

###########################
# models
###########################
# encoders
hidden_size: 192
ffn_hidden_size: 768
enc_ffn_kernel_size: 5
enc_layers: 4
dur_level: word
encoder_type: rel_fft
use_word_encoder: true

# mix ling encoder
word_enc_layers: 4
word_encoder_type: rel_fft
use_pitch_embed: false
enc_prenet: true
enc_pre_ln: true
text_encoder_postnet: true
dropout: 0.0
add_word_pos: true

# dur predictor
dur_predictor_layers: 3
dur_predictor_kernel: 5
predictor_dropout: 0.2

## fvae
use_fvae: true
latent_size: 16
fvae_encoder_type: conv
fvae_decoder_type: conv
fvae_enc_dec_hidden: 192
fvae_kernel_size: 5
fvae_enc_n_layers: 8
fvae_dec_n_layers: 4
fvae_strides: 4
fvae_noise_scale: 1.0

# prior flow
use_prior_flow: true
prior_flow_hidden: 64
prior_flow_kernel_size: 3
prior_flow_n_blocks: 4

###########################
# training and inference
###########################
lambda_kl: 1.0
kl_min: 0.0
lambda_sent_dur: 0.0
kl_start_steps: 10000
posterior_start_steps: 0
frames_multiple: 4
num_valid_plots: 10
lr: 0.0002
warmup_updates: 8000
max_tokens: 40000
valid_infer_interval: 10000
max_sentences: 80
max_updates: 480000