NeuralSeq/configs/tts/fs2.yaml

base_config: configs/tts/base.yaml
task_cls: tasks.tts.fs2.FastSpeech2Task

# model
hidden_size: 256
dropout: 0.1
encoder_type: fft # fft|tacotron|tacotron2|conformer
encoder_K: 8 # for tacotron encoder
decoder_type: fft # fft|rnn|conv|conformer
use_pos_embed: true

# duration
predictor_hidden: -1
predictor_kernel: 5
predictor_layers: 2
dur_predictor_kernel: 3
dur_predictor_layers: 2
predictor_dropout: 0.5

# pitch and energy
use_pitch_embed: true
pitch_type: ph # frame|ph|cwt
use_uv: true
cwt_hidden_size: 128
cwt_layers: 2
cwt_loss: l1
cwt_add_f0_loss: false
cwt_std_scale: 0.8

pitch_ar: false
#pitch_embed_type: 0q
pitch_loss: 'l1' # l1|l2|ssim
pitch_norm: log
use_energy_embed: false

# reference encoder and speaker embedding
use_spk_id: false
use_split_spk_id: false
use_spk_embed: false
use_var_enc: false
lambda_commit: 0.25
ref_norm_layer: bn
pitch_enc_hidden_stride_kernel:
  - 0,2,5 # conv_hidden_size, conv_stride, conv_kernel_size. conv_hidden_size=0: use hidden_size
  - 0,2,5
  - 0,2,5
dur_enc_hidden_stride_kernel:
  - 0,2,3 # conv_hidden_size, conv_stride, conv_kernel_size. conv_hidden_size=0: use hidden_size
  - 0,2,3
  - 0,1,3


# mel
mel_loss: l1:0.5|ssim:0.5 # l1|l2|gdl|ssim or l1:0.5|ssim:0.5

# loss lambda
lambda_f0: 1.0
lambda_uv: 1.0
lambda_energy: 0.1
lambda_ph_dur: 1.0
lambda_sent_dur: 1.0
lambda_word_dur: 1.0
predictor_grad: 0.1

# train and eval
pretrain_fs_ckpt: ''
warmup_updates: 2000
max_tokens: 32000
max_sentences: 100000
max_eval_sentences: 1
max_updates: 120000
num_valid_plots: 5
num_test_samples: 0
test_ids: []
use_gt_dur: false
use_gt_f0: false

# exp
dur_loss: mse # huber|mol
norm_type: gn