config.yaml

encoder: chunkformer
is_json_cmvn: true
cmvn_file: chunkformer-large-vietnamese/global_cmvn
input_dim: 80
output_dim: 6992

encoder_conf:
    output_size: 512    # dimension of attention
    attention_heads: 8
    linear_units: 2048  # the number of units of position-wise feed forward
    num_blocks: 17      # the number of encoder blocks
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    attention_dropout_rate: 0.1
    input_layer: 'depthwise' # encoder input type, you can chose conv2d, conv2d6 and conv2d8
    normalize_before: true
    cnn_module_kernel: 15
    use_cnn_module: true
    activation_type: 'swish'
    pos_enc_layer_type: 'stream_rel_pos'
    selfattention_layer_type: 'stream_rel_selfattn'
    causal: false
    use_dynamic_chunk: false
    use_limited_chunk: false
    use_context_hint_chunk: false
    right_context_probs: [0.75]
    right_context_sizes: [128, 128, 128]
    limited_decoding_chunk_sizes: [64, 128, 256]
    limited_left_chunk_sizes: [128, 256, 128]
    cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster
    use_dynamic_left_chunk: false
    use_dynamic_conv: true
    freeze_subsampling_layer: false