hyperparams.yaml

# ################################
# Model: wav2vec2 + DNN + CTC
# Augmentation: SpecAugment
# Authors: Sung-Lin Yeh 2021, Titouan Parcollet 2022
# ################################

# URL for the biggest Fairseq english wav2vec2 model.
wav2vec2_hub: facebook/wav2vec2-large-960h-lv60-self
sample_rate: 16000

# Model parameters
activation: !name:torch.nn.LeakyReLU
dnn_layers: 2
dnn_neurons: 1024
freeze_wav2vec: True

# Outputs
output_neurons: 31  # BPE size, index(blank/eos/bos) = 0

# Decoding parameters
blank_index: 0
bos_index: 1
eos_index: 2

enc: !new:speechbrain.lobes.models.VanillaNN.VanillaNN
   input_shape: [null, null, 1024]
   activation: !ref <activation>
   dnn_blocks: !ref <dnn_layers>
   dnn_neurons: !ref <dnn_neurons>

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

ctc_lin: !new:speechbrain.nnet.linear.Linear
   input_size: !ref <dnn_neurons>
   n_neurons: !ref <output_neurons>

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

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

asr_model: !new:torch.nn.ModuleList
    - [!ref <enc>, !ref <ctc_lin>]

tokenizer: !new:speechbrain.dataio.encoder.CTCTextEncoder

encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
    wav2vec2: !ref <wav2vec2>
    enc: !ref <enc>
    ctc_lin: !ref <ctc_lin>

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:
      wav2vec2: !ref <wav2vec2>
      asr: !ref <asr_model>
      tokenizer: !ref <tokenizer>