hyperparams.yaml

# ############################################################################
# Model: E2E ST JA->EN with Conformer
# Encoder: Conformer Encoder
# Decoder: Conformer Decoder + (CTC/ATT joint)
# Tokens: BPE
# losses: CTC
# Training: Custom JA->EN youtube scrape, ~600h
# Authors:  Eric Engelhart, 2022
# ############################################################################

# Tokenier initialization
tokenizer: !new:sentencepiece.SentencePieceProcessor

# Features
sample_rate: 16000
n_fft: 400
n_mels: 80

# normalization
normalizer: !new:speechbrain.processing.features.InputNormalization
    norm_type: global

compute_features: !new:speechbrain.lobes.features.Fbank
    sample_rate: !ref <sample_rate>
    n_fft: !ref <n_fft>
    n_mels: !ref <n_mels>

####################### Model parameters ###########################
# Transformer
d_model: 384
nhead: 6
num_encoder_layers: 12
num_decoder_layers: 6
d_ffn: 1536
transformer_dropout: 0.1
activation: !name:torch.nn.GELU
output_neurons: 5000
vocab_size: 5000
attention_type: "regularMHA" # "RelPosMHAXL" or "regularMHA"
kernel_size: 15
encoder_module: conformer


# 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: 2
valid_beam_size: 1
test_beam_size: 25

############################## models ################################
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
    input_shape: (8, 10, 80)
    num_blocks: 2
    num_layers_per_block: 1
    out_channels: (256, 256)
    kernel_sizes: (3, 3)
    strides: (2, 2)
    residuals: (False, False)

Transformer: !new:speechbrain.lobes.models.transformer.TransformerST.TransformerST # yamllint disable-line rule:line-length
    input_size: 5120
    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>
    ctc_weight: 0
    asr_weight: 0
    mt_weight: 0
    asr_tgt_vocab: !ref <output_neurons>
    mt_src_vocab: !ref <output_neurons>
    attention_type: !ref <attention_type>
    kernel_size: !ref <kernel_size>
    encoder_module: !ref <encoder_module>
    normalize_before: True
    causal: False
    max_length: 5000

# only when multi-task setting is used
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>

# when asr-weight > 0 and ctc-weight < 1
asr_seq_lin: !new:speechbrain.nnet.linear.Linear
    input_size: !ref <d_model>
    n_neurons: !ref <vocab_size>


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


Tencoder: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
    transformer: !ref <Transformer>

encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
    input_shape: [null, null, !ref <n_mels>]
    compute_features: !ref <compute_features>
    normalize: !ref <normalizer>
    cnn: !ref <CNN>
    transformer_encoder: !ref <Tencoder>



decoder: !new:speechbrain.decoders.S2STransformerBeamSearch
    modules: [!ref <Transformer>, !ref <seq_lin>, null]
    bos_index: !ref <bos_index>
    eos_index: !ref <eos_index>
    blank_index: !ref <blank_index>
    min_decode_ratio: !ref <min_decode_ratio>
    max_decode_ratio: !ref <max_decode_ratio>
    beam_size: !ref <test_beam_size>
    using_eos_threshold: True
    length_normalization: True
    ctc_weight: 0
    lm_weight: 0


modules:
    compute_features: !ref <compute_features>
    normalizer: !ref <normalizer>
    pre_transformer: !ref <CNN>
    Transformer: !ref <Transformer>
    asr_model: !ref <st_model>
    encoder: !ref <encoder>
    decoder: !ref <decoder>


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



# The pretrainer allows a mapping between pretrained files and instances that
# are declared in the yaml.
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
    loadables:
        tokenizer: !ref <tokenizer>
        st: !ref <st_model>
        normalizer: !ref <normalizer>