model_conf:
  fastemit_lambda: Regularization parameter for FastEmit. (float, default = 0.0)

model_conf:
  use_k2_pruned_loss: True
  k2_pruned_loss_args:
    prune_range: How many tokens by frame are used compute the pruned loss. (int, default = 5)
    simple_loss_scaling: The weight to scale the simple loss after warm-up. (float, default = 0.5)
    lm_scale: The scale factor to smooth the LM part. (float, default = 0.0)
    am_scale: The scale factor to smooth the AM part. (float, default = 0.0)
    loss_type: Define the type of path to take for loss computation, either 'regular', 'smoothed' or 'constrained'. (str, default = "regular")

model_conf:
  validation_nstep: Maximum number of symbol expansions at each time step when reporting CER or/and WER using mAES.

main_conf:
  pos_wise_act_type: Conformer position-wise feed-forward activation type. (str, default = "swish")
  conv_mod_act_type: Conformer convolution module activation type. (str, default = "swish")
  pos_enc_dropout_rate: Dropout rate for the positional encoding layer, if used. (float, default = 0.0)
  pos_enc_max_len: Positional encoding maximum length. (int, default = 5000)
  simplified_att_score: Whether to use simplified attention score computation. (bool, default = False)
  norm_type: X-former normalization module type. (str, default = "layer_norm")
  conv_mod_norm_type: Branchformer convolution module normalization type. (str, default = "layer_norm")
  after_norm_eps: Epsilon value for the final normalization module. (float, default = 1e-05 or 0.25 for BasicNorm)
  after_norm_partial: Partial value for the final normalization module, if norm_type = 'rms_norm'. (float, default = -1.0)
  blockdrop_rate: Probability threshold of dropping out each encoder block. (float, default = 0.0)
  # For more information on the parameters below, please refer to espnet2/asr_transducer/activation.py
  ftswish_threshold: Threshold value for FTSwish activation formulation.
  ftswish_mean_shift: Mean shifting value for FTSwish activation formulation.
  hardtanh_min_val: Minimum value of the linear region range for HardTanh activation. (float, default = -1.0)
  hardtanh_max_val: Maximum value of the linear region range for HardTanh. (float, default = 1.0)
  leakyrelu_neg_slope: Negative slope value for LeakyReLU activation formulation.
  smish_alpha: Alpha value for Smish variant activation fomulation. (float, default = 1.0)
  smish_beta: Beta value for Smish variant activation formulation. (float, default = 1.0)
  softplus_beta: Beta value for softplus activation formulation in Mish activation. (float, default = 1.0)
  softplus_threshold: Values above this revert to a linear function in Mish activation. (int, default = 20)
  swish_beta: Beta value for E-Swish activation formulation. (float, default = 20)

input_conf:
  block_type: Input block type, either "conv2d" or "vgg". (str, default = "conv2d")
  conv_size: Convolution output size. For "vgg", the two convolution outputs can be controlled by passing a tuple. (int, default = 256)
  subsampling_factor: Subsampling factor of the input block, either 2 (only conv2d), 4 or 6. (int, default = 4)

    # Branchformer
    - block_type: branchformer
      hidden_size: Hidden (and output) dimension. (int)
      linear_size: Dimension of the Linear layers. (int)
      conv_mod_kernel_size: Size of the convolving kernel in the ConvolutionalSpatialGatingUnit module. (int)
      heads (optional): Number of heads in multi-head attention. (int, default = 4)
      norm_eps (optional): Epsilon value for the normalization module. (float, default = 1e-05 or 0.25 for BasicNorm)
      norm_partial (optional): Partial value for the normalization module, if norm_type = 'rms_norm'. (float, default = -1.0)
      conv_mod_norm_eps (optional): Epsilon value for ConvolutionalSpatialGatingUnit module normalization. (float, default = 1e-05 or 0.25 for BasicNorm)
      conv_mod_norm_partial (optional): Partial value for the ConvolutionalSpatialGatingUnit module normalization, if conv_norm_type = 'rms_norm'. (float, default = -1.0)
      dropout_rate (optional): Dropout rate for some intermediate layers. (float, default = 0.0)
      att_dropout_rate (optional): Dropout rate for the attention module. (float, default = 0.0)

    # Conformer
    - block_type: conformer
      hidden_size: Hidden (and output) dimension. (int)
      linear_size: Dimension of feed-forward module. (int)
      conv_mod_kernel_size: Size of the convolving kernel in the ConformerConvolution module. (int)
      heads (optional): Number of heads in multi-head attention. (int, default = 4)
      norm_eps (optional): Epsilon value for normalization module. (float, default = 1e-05 or 0.25 for BasicNorm)
      norm_partial (optional): Partial value for the normalization module, if norm_type = 'rms_norm'. (float, default = -1.0)
      conv_mod_norm_eps (optional): Epsilon value for Batchnorm1d in the ConformerConvolution module. (float, default = 1e-05)
      conv_mod_norm_momentum (optional): Momentum value for Batchnorm1d in the ConformerConvolution module. (float, default = 0.1)
      dropout_rate (optional): Dropout rate for some intermediate layers. (float, default = 0.0)
      att_dropout_rate (optional): Dropout rate for the attention module. (float, default = 0.0)
      pos_wise_dropout_rate (optional): Dropout rate for the position-wise feed-forward module. (float, default = 0.0)

    # Conv 1D
    - block_type: conv1d
      output_size: Output size. (int)
      kernel_size: Size of the convolving kernel. (int or Tuple)
      stride (optional): Stride of the sliding blocks. (int or tuple, default = 1)
      dilation (optional): Parameter to control the stride of elements within the neighborhood. (int or tuple, default = 1)
      groups (optional): Number of blocked connections from input channels to output channels. (int, default = 1)
      bias (optional): Whether to add a learnable bias to the output. (bool, default = True)
      relu (optional): Whether to use a ReLU activation after convolution. (bool, default = True)
      batch_norm: Whether to use batch normalization after convolution. (bool, default = False)
      dropout_rate (optional): Dropout rate for the Conv1d outputs. (float, default = 0.0)

    # E-Branchformer
    - block_type: ebranchformer
      hidden_size: Hidden (and output) dimension. (int)
      linear_size: Dimension of the feed-forward module and othger linear layers. (int)
      conv_mod_kernel_size: Size of the convolving kernel in the ConvolutionalSpatialGatingUnit module. (int)
      depthwise_conv_kernel_size: Size of the convolving kernel in the DepthwiseConvolution module. (int, default = conv_mod_kernel_size)
      heads (optional): Number of heads in multi-head attention. (int, default = 4)
      norm_eps (optional): Epsilon value for the normalization module. (float, default = 1e-05 or 0.25 for BasicNorm)
      norm_partial (optional): Partial value for the normalization module, if norm_type = 'rms_norm'. (float, default = -1.0)
      conv_mod_norm_eps (optional): Epsilon value for ConvolutionalSpatialGatingUnit module normalization. (float, default = 1e-05 or 0.25 for BasicNorm)
      conv_mod_norm_partial (optional): Partial value for the ConvolutionalSpatialGatingUnit module normalization, if conv_norm_type = 'rms_norm'. (float, default = -1.0)
      dropout_rate (optional): Dropout rate for some intermediate layers. (float, default = 0.0)
      att_dropout_rate (optional): Dropout rate for the attention module. (float, default = 0.0)

encoder_conf:
    main_conf:
      pos_wise_act_type: swish
      pos_enc_dropout_rate: 0.1
      conv_mod_act_type: swish
    input_conf:
      block_type: conv2d
      conv_size: 256
      subsampling_factor: 4
    body_conf:
    - block_type: conv1d
      output_size: 128
      kernel_size: 3
    - block_type: conv1d
      output_size: 256
      kernel_size: 2
    - block_type: conformer
      linear_size: 1024
      hidden_size: 256
      heads: 8
      dropout_rate: 0.1
      pos_wise_dropout_rate: 0.1
      att_dropout_rate: 0.1
      conv_mod_kernel_size: 31
      num_blocks: 14

decoder_conf:
  embed_size: Size of the embedding layer (int, default = 256).
  num_blocks: Number of decoder blocks/layers (int, default = 4 for MEGA or 1 for RNN).
  rnn_type (RNN only): Type of RNN cells (int, default = "lstm").
  hidden_size (RNN only): Size of the hidden layers (int, default = 256).
  block_size (MEGA/RWKV only): Size of the block's input/output (int, default = 512).
  linear_size (MEGA/RWKV only): Feed-Forward module hidden size (int, default = 1024).
  attention_size (RWKV only): Hidden-size of the attention module. (int, default = None).
  context_size (RWKV only): Context size for the WKV kernel module (int, default = 1024).
  qk_size (MEGA only): Shared query and key size for attention module (int, default = 128).
  v_size (MEGA only): Value size for attention module (int, default = 1024).
  chunk_size (MEGA only): Chunk size for attention computation (int, default = -1, i.e. full context).
  num_heads (MEGA only): Number of EMA heads (int, default = 4).
  rel_pos_bias (MEGA only): Type of relative position bias in attention module (str, default = "simple").
  max_positions (MEGA only): Maximum number of position for RelativePositionBias (int, default = 2048).
  truncation_length (MEGA only): Maximum length for truncation in EMA module (int, default = None).
  normalization_type (MEGA/RWKV only): Normalization layer type (str, default = "layer_norm").
  normalization_args (MEGA/RKWV only): Normalization layer arguments (dict, default = {}).
  activation_type (MEGA only): Activation function type (str, default = "swish").
  activation_args (MEGA only): Activation function arguments (dict, default = {}).
  rescale_every (RWKV only): Whether to rescale input every N blocks during inference (int, default = 0)
  dropout_rate (excl. RWKV): Dropout rate for main block modules (float, default = 0.0).
  embed_dropout_rate: Dropout rate for embedding layer (float, default = 0.0).
  att_dropout_rate (MEGA/RWKV only): Dropout rate for the attention module.
  ema_dropout_rate (MEGA only): Dropout rate for the EMA module.
  ffn_dropout_rate (MEGA/RWKV only): Dropout rate for the feed-forward module.

decoder: rnn
decoder_conf:
    rnn_type: lstm
    num_layers: 2
    embed_size: 256
    hidden_size: 256
    dropout_rate: 0.1
    embed_dropout_rate: 0.1

decoder: mega
decoder_conf:
    block_size: 256
    linear_size: 2048
    qk_size: 128
    v_size: 1024
    max_positions: 1024
    num_heads: 4
    rel_pos_bias_type: "rotary"
    chunk_size: 256
    num_blocks: 6
    dropout_rate: 0.1
    ffn_dropout_rate: 0.1
    att_dropout_rate: 0.1
    embed_dropout_rate: 0.1

model_conf:
  transducer_weight: Weight of the Transducer loss (float, default = 1.0)
  auxiliary_ctc_weight: Weight of the CTC loss. (float, default = 0.0)
  auxiliary_ctc_dropout_rate: Dropout rate for the CTC loss inputs. (float, default = 0.0)
  auxiliary_lm_loss_weight: Weight of the LM loss. (float, default = 0.2)
  auxiliary_lm_loss_smoothing: Smoothing rate for LM loss. If > 0, label smoothing is enabled. (float, default = 0.0)

search_type: tsd
max_sym_exp : Number of maximum symbol expansions at each time step. (int > 1, default = 3)

search_type: alsd
u_max: Maximum expected target sequence length. (int, default = 50)

search_type: maes
nstep: Number of maximum expansion steps at each time step (int, default = 2)
expansion_gamma: Number of additional candidates in expanded hypotheses selection. (int, default = 2)
expansion_beta: Allowed logp difference for prune-by-value method. (float, default = 2.3)

dynamic_chunk_training: Whether to train streaming model with dynamic chunks. (bool, default = False)
short_chunk_threshold: Chunk length threshold (in percent) for dynamic chunk selection. (int, default = 0.75)
short_chunk_size: Minimum number of frames during dynamic chunk training. (int, default = 25)
num_left_chunks: The number of left chunks the attention module can see during training, where the actual size is defined by `short_chunk_threshold` and `short_chunk_size`. (int, default = 0, i.e. full context)

decoding_window: The input audio length, in milliseconds, to process during decoding. (int, default = 640)
left_context: Number of previous frames (AFTER subsampling) the attention module can see in current chunk. (int, default = 32)