sample_rate: 16000
timesteps: 64
repeat: 2
dropout: 0.0
kernel_size_factor: 1.0
labels:
- background
- speech
train_ds:
  vad_stream: false
  manifest_filepath: /home/fjia/code/manifest64/train.json
  sample_rate: 16000
  labels:
  - background
  - speech
  batch_size: 128
  num_workers: 20
  shuffle: true
  augmentor:
    shift:
      prob: 0.8
      min_shift_ms: -5.0
      max_shift_ms: 5.0
    white_noise:
      prob: 0.8
      min_level: -90
      max_level: -46
validation_ds:
  vad_stream: false
  manifest_filepath: /home/fjia/code/manifest64/validation.json
  sample_rate: 16000
  labels:
  - background
  - speech
  batch_size: 128
  shuffle: false
  val_loss_idx: 0
  num_workers: 20
test_ds:
  vad_stream: false
  manifest_filepath: null
  sample_rate: 16000
  labels:
  - background
  - speech
  batch_size: 128
  shuffle: false
  test_loss_idx: 0
  num_workers: 20
preprocessor:
  cls: nemo.collections.asr.modules.AudioToMFCCPreprocessor
  params:
    window_size: 0.025
    window_stride: 0.01
    window: hann
    n_mels: 64
    n_mfcc: 64
    n_fft: 512
spec_augment:
  cls: nemo.collections.asr.modules.SpectrogramAugmentation
  params:
    freq_masks: 2
    time_masks: 2
    freq_width: 15
    time_width: 25
    rect_masks: 5
    rect_time: 25
    rect_freq: 15
encoder:
  cls: nemo.collections.asr.modules.ConvASREncoder
  params:
    feat_in: 64
    activation: relu
    conv_mask: true
    jasper:
    - filters: 128
      repeat: 1
      kernel:
      - 11
      stride:
      - 1
      dilation:
      - 1
      dropout: 0.0
      residual: false
      separable: true
      kernel_size_factor: 1.0
    - filters: 64
      repeat: 2
      kernel:
      - 13
      stride:
      - 1
      dilation:
      - 1
      dropout: 0.0
      residual: true
      separable: true
      kernel_size_factor: 1.0
    - filters: 64
      repeat: 2
      kernel:
      - 15
      stride:
      - 1
      dilation:
      - 1
      dropout: 0.0
      residual: true
      separable: true
      kernel_size_factor: 1.0
    - filters: 64
      repeat: 2
      kernel:
      - 17
      stride:
      - 1
      dilation:
      - 1
      dropout: 0.0
      residual: true
      separable: true
      kernel_size_factor: 1.0
    - filters: 128
      repeat: 1
      kernel:
      - 29
      stride:
      - 1
      dilation:
      - 2
      dropout: 0.0
      residual: false
      separable: true
      kernel_size_factor: 1.0
    - filters: 128
      repeat: 1
      kernel:
      - 1
      stride:
      - 1
      dilation:
      - 1
      dropout: 0.0
      residual: false
decoder:
  cls: nemo.collections.asr.modules.ConvASRDecoderClassification
  params:
    feat_in: 128
    num_classes: 2
    return_logits: true
    pooling_type: avg
optim:
  name: sgd
  lr: 0.01
  weight_decay: 0.001
  sched:
    name: PolynomialHoldDecayAnnealing
    power: 2.0
    warmup_ratio: 0.05
    hold_ratio: 0.45
    min_lr: 0.001
    last_epoch: -1
  momentum: 0.9
target: nemo.collections.asr.models.classification_models.EncDecClassificationModel