train/models/s2t/medium_config.json

{
  "emb_size": 512,
  "feedforward_size": 2048,
  "hidden_size": 512,
  "hidden_act": "relu",
  "heads_num": 8,
  "layers_num": 12,
  "decoder_layers_num": 6,
  "max_audio_frames": 6000,
  "dropout": 0.1,
  "data_processor": "s2t",
  "embedding": ["speech", "sinusoidalpos"],
  "tgt_embedding": ["word", "sinusoidalpos"],
  "encoder": "transformer",
  "mask": "fully_visible",
  "decoder": "transformer",
  "target": ["lm"],
  "has_lmtarget_bias": false,
  "conv_channels": [1024, 1024],
  "audio_feature_size": 80,
  "conv_kernel_sizes": [5, 5],
  "layernorm_positioning": "pre",
  "remove_embedding_layernorm": true,
  "tie_weights": true,
  "optimizer": "adamw",
  "scheduler": "inverse_sqrt",
  "audio_preprocess": ["normalize_means", "normalize_vars", "ceptral_normalize"],
  "specaugment":{
    "freq_mask_F": 27,
    "freq_mask_N": 2,
    "time_mask_N": 2,
    "time_mask_T": 100,
    "time_mask_p": 1.0,
    "time_wrap_W": 0
   },
   "label_smoothing": 0.1,
   "ignore_index": true
}