hyperparams.yaml

# ############################################################################
# Model: E2E ASR with Transformer + wav2vec2
# Encoder: wav2vec Encoder
# Decoder: Transformer Decoder + (CTC/ATT joint) beamsearch
# Tokens: BPE with unigram
# losses: CTC + KLdiv (Label Smoothing loss)
# Training: AISHELL-1
# Authors:  Jianyuan Zhong, Titouan Parcollet, Mirco Ravanelli
# ############################################################################

# Feature parameters
sample_rate: 16000
n_fft: 400
n_mels: 80
wav2vec2_hub: facebook/wav2vec2-large-100k-voxpopuli

####################### Model parameters ###########################
# Transformer
d_model: 256
nhead: 4
num_encoder_layers: 2
num_decoder_layers: 6
d_ffn: 2048
transformer_dropout: 0.1
activation: !name:torch.nn.GELU
output_neurons: 5000
vocab_size: 5000

# Outputs
blank_index: 0
label_smoothing: 0.1
pad_index: 0
bos_index: 1
eos_index: 2
unk_index: 0

# Decoding parameters
min_decode_ratio: 0.0
max_decode_ratio: 1.0
valid_search_interval: 10
valid_beam_size: 10
test_beam_size: 10
ctc_weight_decode: 0.40

############################## models ################################

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

Transformer: !new:speechbrain.lobes.models.transformer.TransformerASR.TransformerASR # yamllint disable-line rule:line-length
    input_size: 1024
    tgt_vocab: !ref <output_neurons>
    d_model: !ref <d_model>
    nhead: !ref <nhead>
    num_encoder_layers: !ref <num_encoder_layers>
    num_decoder_layers: !ref <num_decoder_layers>
    d_ffn: !ref <d_ffn>
    dropout: !ref <transformer_dropout>
    activation: !ref <activation>
    normalize_before: True


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

seq_lin: !new:speechbrain.nnet.linear.Linear
    input_size: !ref <d_model>
    n_neurons: !ref <output_neurons>

tokenizer: !new:sentencepiece.SentencePieceProcessor

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

# Here, we extract the encoder from the Transformer model
Tencoder: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
    transformer: !ref <Transformer>

# We compose the inference (encoder) pipeline.
encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
    input_shape: [null, null, !ref <n_mels>]
    wav2vec2: !ref <wav2vec2>
    transformer_encoder: !ref <Tencoder>

ctc_scorer: !new:speechbrain.decoders.scorer.CTCScorer
    eos_index: !ref <eos_index>
    blank_index: !ref <blank_index>
    ctc_fc: !ref <ctc_lin>
    
scorer: !new:speechbrain.decoders.scorer.ScorerBuilder
    full_scorers: [!ref <ctc_scorer>]
    weights:
        ctc: !ref <ctc_weight_decode>
        
decoder: !new:speechbrain.decoders.S2STransformerBeamSearcher
    modules: [!ref <Transformer>, !ref <seq_lin>]
    bos_index: !ref <bos_index>
    eos_index: !ref <eos_index>
    min_decode_ratio: !ref <min_decode_ratio>
    max_decode_ratio: !ref <max_decode_ratio>
    beam_size: !ref <test_beam_size>
    using_eos_threshold: False
    length_normalization: True
    scorer: !ref <scorer>

modules:
    encoder: !ref <encoder>
    decoder: !ref <decoder>

log_softmax: !new:torch.nn.LogSoftmax
    dim: -1

pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
    loadables:
        wav2vec2: !ref <wav2vec2>
        model: !ref <asr_model>
        tokenizer: !ref <tokenizer>