asr_example.yaml

data:
  corpus:                                 
    name: 'Librispeech'                   # Specify corpus
    path: 'data/LibriSpeech'          # Path to raw LibriSpeech dataset
    train_split: ['train-clean-100','train-clean-360'] # Name of data splits to be used as training set
    dev_split: ['dev-clean']              # Name of data splits to be used as validation set
    bucketing: True                       # Enable/Disable bucketing 
    batch_size: 16
  audio:                                  # Attributes of audio feature
    feat_type: 'fbank'
    feat_dim:  40
    frame_length: 25                      # ms
    frame_shift: 10                       # ms
    dither: 0                             # random dither audio, 0: no dither
    apply_cmvn: True
    delta_order: 2                        # 0: do nothing, 1: add delta, 2: add delta and accelerate
    delta_window_size: 2
  text:
    mode: 'subword'                       # 'character'/'word'/'subword'
    vocab_file: 'tests/sample_data/subword-16k.model'

hparas:                                   # Experiment hyper-parameters
  valid_step: 5000
  max_step: 1000001
  tf_start: 1.0
  tf_end: 1.0
  tf_step: 500000
  optimizer: 'Adadelta'
  lr: 1.0
  eps: 0.00000001                         # 1e-8
  lr_scheduler: 'fixed'                   # 'fixed'/'warmup'
  curriculum: 0

model:                                    # Model architecture
  ctc_weight: 0.0                         # Weight for CTC loss
  encoder:
    prenet: 'vgg'                         # 'vgg'/'cnn'/''
    # vgg: True                             # 4x reduction on time feature extraction
    module: 'LSTM'                        # 'LSTM'/'GRU'/'Transformer'
    bidirection: True
    dim: [512,512,512,512,512]
    dropout: [0,0,0,0,0]
    layer_norm: [False,False,False,False,False]
    proj: [True,True,True,True,True]      # Linear projection + Tanh after each rnn layer
    sample_rate: [1,1,1,1,1]
    sample_style: 'drop'                  # 'drop'/'concat'
  attention:
    mode: 'loc'                           # 'dot'/'loc'
    dim: 300
    num_head: 1
    v_proj: False                         # if False and num_head>1, encoder state will be duplicated for each head
    temperature: 0.5                      # scaling factor for attention
    loc_kernel_size: 100                  # just for mode=='loc'
    loc_kernel_num: 10                    # just for mode=='loc'
  decoder:
    module: 'LSTM'                        # 'LSTM'/'GRU'/'Transformer'
    dim: 512
    layer: 1
    dropout: 0