model:
  name: "metro-small"
  d_model: 384
  n_layers: 12
  n_heads: 6
  ff_multiplier: 3
  conv_kernel_size: 31
  dropout: 0.1
  conv_expansion_factor: 2
  se_ratio: 4
  stochastic_depth_rate: 0.05

audio:
  sample_rate: 16000
  n_mels: 80
  n_fft: 512
  hop_length: 160
  win_length: 400

tokenizer:
  type: "bpe"
  vocab_size: 5000

training:
  intermediate_ctc_layers: [6]