asr_training_args.yml

entry.class: trainer
entry.params:
  train_steps: 200000
  summary_steps: 200
  save_checkpoint_steps: 2000
  criterion.class: label_smoothed_cross_entropy
  criterion.params:
    label_smoothing: 0.1
  optimizer.class: adam
  optimizer.params:
    epsilon: 1.e-9
    beta_1: 0.9
    beta_2: 0.98
  lr_schedule.class: noam
  lr_schedule.params:
    initial_factor: 3.5
    end_factor: 2.0
    dmodel: 256
    warmup_steps: 25000
    start_decay_at: 50000
    decay_steps: 50000

dataset.class: AudioTFRecordDataset
dataset.params:
  data_path: DATA_PATH/asr_st/train/
  shuffle_dataset: True
  feature_key: audio
  transcript_key: transcript

task.class: SpeechToText
task.params:
  audio_feature_dim: 80
  audio_feature_channels: 1
  transcript_data_pipeline.class: TranscriptDataPipeline
  transcript_data_pipeline.params:
    remove_punctuation: True
    lowercase: True
    language: en
    tokenizer: moses
    subtokenizer: bpe
    subtokenizer_codes: DATA_PATH/asr_st/codes.bpe
    vocab_path: DATA_PATH/asr_st/vocab.en
  batch_by_frames: True
  batch_size: 120000
  max_src_len: 3000
  max_trg_len: 120
  truncate_src: True
  experimental_frame_transcript_ratio: 21