model_config.yaml

decoder:
  _target_: nemo.collections.asr.modules.SpeakerDecoder
  angular: false
  emb_sizes: 192
  feat_in: 3072
  num_classes: 2
  pool_mode: attention
encoder:
  _target_: nemo.collections.asr.modules.ConvASREncoder
  activation: relu
  conv_mask: true
  feat_in: 80
  jasper:
  - dilation:
    - 1
    dropout: 0.0
    filters: 1024
    kernel:
    - 3
    repeat: 1
    residual: false
    se: true
    se_context_size: -1
    separable: true
    stride:
    - 1
  - dilation:
    - 1
    dropout: 0.1
    filters: 1024
    kernel:
    - 7
    repeat: 3
    residual: true
    se: true
    se_context_size: -1
    separable: true
    stride:
    - 1
  - dilation:
    - 1
    dropout: 0.1
    filters: 1024
    kernel:
    - 11
    repeat: 3
    residual: true
    se: true
    se_context_size: -1
    separable: true
    stride:
    - 1
  - dilation:
    - 1
    dropout: 0.1
    filters: 1024
    kernel:
    - 15
    repeat: 3
    residual: true
    se: true
    se_context_size: -1
    separable: true
    stride:
    - 1
  - dilation:
    - 1
    dropout: 0.0
    filters: 3072
    kernel:
    - 1
    repeat: 1
    residual: false
    se: true
    se_context_size: -1
    separable: true
    stride:
    - 1
loss:
  margin: 0.2
  scale: 30
model_defaults:
  dropout: 0.1
  enc_hidden: 640
  filters: 1024
  joint_hidden: 640
  kernel_size_factor: 1.0
  pred_hidden: 640
  repeat: 3
  se: true
  se_context_size: -1
  separable: true
optim:
  lr: 0.08
  momentum: 0.9
  name: sgd
  sched:
    min_lr: 0.0
    name: CosineAnnealing
    warmup_ratio: 0.1
  weight_decay: 0.0002
preprocessor:
  _target_: nemo.collections.asr.modules.AudioToMelSpectrogramPreprocessor
  dither: 1.0e-05
  features: 80
  frame_splicing: 1
  n_fft: 512
  normalize: per_feature
  sample_rate: 16000
  window: hann
  window_size: 0.025
  window_stride: 0.01
spec_augment:
  _target_: nemo.collections.asr.modules.SpectrogramAugmentation
  freq_masks: 3
  freq_width: 4
  time_masks: 5
  time_width: 0.03
target: nemo.collections.asr.models.label_models.EncDecSpeakerLabelModel
train_ds:
  augmentor:
    noise:
      manifest_path: /manifests/noise/rir_noise_manifest.json
      max_snr_db: 15
      min_snr_db: 0
      prob: 0.5
    speed:
      max_speed_rate: 1.05
      min_speed_rate: 0.95
      prob: 0.5
      resample_type: kaiser_fast
      sr: 16000
  batch_size: 64
  is_tarred: false
  labels: null
  manifest_filepath: /manifests/combined_fisher_swbd_voxceleb12_librispeech/train.json
  num_workers: 15
  pin_memory: true
  sample_rate: 16000
  shuffle: true
  tarred_audio_filepaths: null
  tarred_shard_strategy: scatter
  time_length: 3
validation_ds:
  batch_size: 128
  labels: null
  manifest_filepath: /manifests/combined_fisher_swbd_voxceleb12_librispeech/dev.json
  num_workers: 15
  pin_memory: true
  sample_rate: 16000
  shuffle: false
  time_length: 3