default.yaml

data:
  datasets_path: './datasets'
  dataset: 'ljspeech'
  dataset_dir: 'LJSpeech-1.1'
text:
  graphemes: &gs !!python/object/apply:eval ['list("abcdefghijklmnopqrstuvwxyz")']
  phonemes: &ps !!python/object/apply:eval ['["AA0","AA1","AA2","AE0","AE1","AE2","AH0","AH1","AH2","AO0","AO1","AO2","AW0","AW1","AW2","AY0","AY1","AY2","B","CH","D","DH","EH0","EH1","EH2","ER0","ER1","ER2","EY0","EY1","EY2","F","G","HH","IH0","IH1","IH2","IY0","IY1","IY2","JH","K","L","M","N","NG","OW0","OW1","OW2","OY0","OY1","OY2","P","R","S","SH","T","TH","UH0","UH1","UH2","UW","UW0","UW1","UW2","V","W","Y","Z","ZH"]']
  specials: &sp !!python/object/apply:eval ['["<pad>", "<unk>"]']
  punctuations: &pt !!python/object/apply:eval ['[".", ",", "?", "!", " ", "-"]']
  units_list: &ul !!python/object/apply:eval ['us+sp+pt', {'us': *ps, 'sp': *sp, 'pt': *pt}]
  use_phonemes: &up true
audio:
  n_mel_channels: &nm 80
  filter_length: 1024
  hop_length: 256  # WARNING: this can't be changed.
  win_length: 1024
  sampling_rate: &sr 22050
  segment_length: *sr
  pad_short: 2000
  mel_fmin: 0.0
  mel_fmax: 8000.0
  # Precomputed statistics for log-mel-spectrs for speech dataset
  spec_mean: -5.522  # for LJSpeech dataset
  spec_std: 2.063  # for LJSpeech dataset
  spec_min: -11.5129  # for LJSpeech dataset
  spec_max: 2.0584  # for LJSpeech dataset
  # Others
  force_frame_rate: true  # force match sampling rate
  normalize: 
    match_volume: false
    trim_silence: false
  reduction_rate: 4
parallel:
  ground_truth: false
  out_channels: *nm  # equal to ${audio.n_mel_channels}
  alphabet_size: !!python/object/apply:eval ['len(ul)', {'ul': *ul}]
  channels: 128
  enc_kernel_size: 4
  dec_kernel_size: 4
  enc_dilations: !!python/object/apply:eval ['4 * [1,2,4]   + [1]']  # receptive field is max 15
  dec_dilations: !!python/object/apply:eval ['4 * [1,2,4,8] + [1]']  # receptive field is max 32
  normalize: FreqNorm  # 'freq', 'layer', 'batch'
  activation: torch.nn.ReLU  # 'relu', 'linear', 'sigmoid'
  final_activation: torch.nn.Identity
  pos_mode: 'duration'  # 'standard', 'duration'
  interpolate: false  # true
  separate_duration_grad: true
  speaker_dim: 128  # speaker embedding dim
duration:
  out_channels: *nm  # equal to ${audio.n_mel_channels}
  alphabet_size: !!python/object/apply:eval ['len(ul)', {'ul': *ul}]
  channels: 40
  hidden_channels: 80
  kernel_size: 3
  text_enc_dilations: !!python/object/apply:eval ['2 * [1,3,9,27] + [1,1]']
  spec_enc_dilations: !!python/object/apply:eval ['2 * [1,3,9,27] + [1,1]']
  spec_dec_dilations: !!python/object/apply:eval ['2 * [1,3,9,27] + [1,1]']
  normalize: torch.nn.Linear  # 'layer'  # 'freq', 'layer', 'batch'
  activation: torch.nn.ReLU  # 'relu', 'linear', 'sigmoid'
  final_activation: torch.nn.Sigmoid
  att_noise: 0.1
  att_hidden_channels: 80
  pos_mode: 'duration'  # 'standard', 'duration'
  # Spectrogram augmentation options
  enable_augment: true
  noise: 0.01  # 1. add normal noise to input spectrs
  feed_repeat: 2  # 2. Feed spectrograms through the model `feed_repeat` times, use degraded output on input for training
  feed_ratio: 0.5  # how many items in batch are degraded
  replace_ratio: 0.1  # 3. Replace random spectrogram frames with random noise
melgan:
  checkpoint: 'melgan-epoch3200.pth'
trainer:
  disable_progress_bar: false
  logdir: './logdir'
synthesizer:
  inputs_file_path: './outputs/text/synthesize.txt'
  outputs_dir: './outputs/audio'