hyperparams.yaml

# ################################
# Model: wav2vec2 + CTC
# Augmentation: SpecAugment
# Authors: Le Do Thanh Binh 2021
# ################################

sample_rate: 16000
wav2vec2_hub: dragonSwing/wav2vec2-base-vn-270h

# Model parameters
activation: !name:torch.nn.LeakyReLU
dnn_layers: 2
dnn_neurons: 768
dropout_prob: 0.1

# Outputs
output_neurons: 696 # BPE size, index(blank/eos/bos) = 0
output_tones: 12

# Decoding parameters
# Be sure that the bos and eos index match with the BPEs ones
blank_index: 0
bos_index: 1
eos_index: 2
unk_index: 3

tokenizer: !apply:custom.Wav2Vec2WordpieceTokenizer.from_pretrained
  pretrained_model_name_or_path: !ref <wav2vec2_hub>

dropout: !new:torch.nn.Dropout
  p: !ref <dropout_prob>

wav2vec2: !new:speechbrain.lobes.models.huggingface_wav2vec.HuggingFaceWav2Vec2
  source: !ref <wav2vec2_hub>
  output_norm: True
  freeze: True
  save_path: model_checkpoints

lm_head: !new:torch.nn.Linear
  in_features: !ref <dnn_neurons>
  out_features: !ref <output_neurons>

tone_head: !new:torch.nn.Linear
  in_features: !ref <dnn_neurons>
  out_features: !ref <output_tones>

log_softmax: !new:speechbrain.nnet.activations.Softmax
  apply_log: True

ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
  blank_index: !ref <blank_index>

model: !new:torch.nn.ModuleList
  - [!ref <wav2vec2>, !ref <dropout>, !ref <lm_head>, !ref <tone_head>]

encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
  wav2vec2: !ref <wav2vec2>
  dropout: !ref <dropout>
  lm_head: !ref <lm_head>

decoding_function: !name:speechbrain.decoders.ctc_greedy_decode
  blank_id: !ref <blank_index>

modules:
  encoder: !ref <encoder>

pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
  loadables:
    model: !ref <model>