hyperparams.yaml

# ################################
# Model: Wav2Vec + DNN + CTC + Softmax
# Authors:
# Gaelle Laperriere 2023
# ################################

wav2vec_url: LeBenchmark/wav2vec2-FR-3K-large

# Feature parameters:
sample_rate: 16000
feats_dim: 1024

# Model parameters:
activation: !name:torch.nn.LeakyReLU
dnn_blocks: 3
dnn_neurons: 512
log_softmax: !new:torch.nn.LogSoftmax
    dim: -1
    
# Decoding parameters:
blank_index: 0

# Outputs:
output_neurons: 212

# ------ Functions and classes

wav2vec2: !new:speechbrain.lobes.models.huggingface_transformers.wav2vec2.Wav2Vec2
    source: !ref <wav2vec_url>
    output_norm: True
    freeze: True
    save_path: wav2vec2_checkpoint

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

output_lin: !new:speechbrain.nnet.linear.Linear
    input_size: !ref <dnn_neurons>
    n_neurons: !ref <output_neurons>
    bias: True

model: !new:torch.nn.ModuleList
    - [!ref <enc>, !ref <output_lin>]
    
tokenizer: !new:speechbrain.dataio.encoder.CTCTextEncoder

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

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>
        model: !ref <model>
        tokenizer: !ref <tokenizer>