examples/asr/conf/quartznet/quartznet_15x5_ru.yaml

name: &name "QuartzNet15x5_ru"

model:
  sample_rate: &sample_rate 16000
  repeat: &repeat 5
  dropout: &dropout 0.0
  separable: &separable true
  labels: &labels [ " ", "а", "б", "в", "г", "д", "е", "ж", "з", "и", "й", "к", "л", "м", "н", "о", "п",
                    "р", "с", "т", "у", "ф", "х", "ц", "ч", "ш", "щ", "ъ", "ы", "ь", "э", "ю", "я" ]

  train_ds:
    manifest_filepath: golos/train/manifest.jsonl # Can be found at https://sc.link/JpD
    sample_rate: 16000
    labels: *labels
    batch_size: 32
    trim_silence: True
    max_duration: 20.0
    shuffle: True
    # tarred datasets
    is_tarred: false
    tarred_audio_filepaths: null
    shuffle_n: 2048
    # bucketing params
    bucketing_strategy: "synced_randomized"
    bucketing_batch_size: null
    parser: ru

  validation_ds:
    manifest_filepath: golos/test/crowd/crowd.jsonl # Can be found at https://sc.link/Kqr 
    sample_rate: 16000
    labels: *labels
    batch_size: 32
    shuffle: False
    parser: ru

  test_ds:
    manifest_filepath: golos/test/farfield/farfield.jsonl # Can be found at https://sc.link/Kqr 
    sample_rate: 16000
    labels: *labels
    batch_size: 32
    shuffle: False
    parser: ru

  preprocessor:
    _target_: nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor
    normalize: "per_feature"
    window_size: 0.02
    sample_rate: *sample_rate
    window_stride: 0.01
    window: "hann"
    features: &n_mels 64
    n_fft: 512
    frame_splicing: 1
    dither: 0.00001

  spec_augment:
    _target_: nemo.collections.asr.modules.SpectrogramAugmentation
    rect_freq: 50
    rect_masks: 2
    rect_time: 120

  encoder:
    _target_: nemo.collections.asr.modules.ConvASREncoder
    feat_in: *n_mels
    activation: relu
    conv_mask: true

    jasper:
    - dilation: [1]
      dropout: *dropout
      filters: 256
      kernel: [33]
      repeat: 1
      residual: false
      separable: *separable
      stride: [2]

    - dilation: [1]
      dropout: *dropout
      filters: 256
      kernel: [33]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 256
      kernel: [33]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 256
      kernel: [33]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 256
      kernel: [39]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 256
      kernel: [39]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 256
      kernel: [39]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 512
      kernel: [51]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 512
      kernel: [51]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 512
      kernel: [51]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 512
      kernel: [63]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 512
      kernel: [63]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 512
      kernel: [63]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 512
      kernel: [75]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 512
      kernel: [75]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: 512
      kernel: [75]
      repeat: *repeat
      residual: true
      separable: *separable
      stride: [1]

    - dilation: [2]
      dropout: *dropout
      filters: 512
      kernel: [87]
      repeat: 1
      residual: false
      separable: *separable
      stride: [1]

    - dilation: [1]
      dropout: *dropout
      filters: &enc_filters 1024
      kernel: [1]
      repeat: 1
      residual: false
      stride: [1]

  decoder:
    _target_: nemo.collections.asr.modules.ConvASRDecoder
    feat_in: *enc_filters
    num_classes: 33
    vocabulary: *labels

  optim:
    name: novograd
    # _target_: nemo.core.optim.optimizers.Novograd
    lr: .05
    # optimizer arguments
    betas: [0.8, 0.5]
    weight_decay: 0.001

    # scheduler setup
    sched:
      name: CosineAnnealing

      # pytorch lightning args
      # monitor: val_loss
      # reduce_on_plateau: false

      # Scheduler params
      warmup_steps: null
      warmup_ratio: null
      min_lr: 0.0
      last_epoch: -1

trainer:
  devices: 1 # number of gpus
  max_epochs: 5
  max_steps: -1 # computed at runtime if not set
  num_nodes: 1
  accelerator: gpu
  strategy: ddp
  accumulate_grad_batches: 1
  enable_checkpointing: False  # Provided by exp_manager
  logger: False  # Provided by exp_manager
  log_every_n_steps: 1  # Interval of logging.
  val_check_interval: 1.0  # Set to 0.25 to check 4 times per epoch, or an int for number of iterations
  benchmark: false # needs to be false for models with variable-length speech input as it slows down training

exp_manager:
  exp_dir: null
  name: *name
  create_tensorboard_logger: True
  create_checkpoint_callback: True
  checkpoint_callback_params:
    monitor: "val_wer"
    mode: "min"
  create_wandb_logger: False
  wandb_logger_kwargs:
    name: null
    project: null