hyperparams.yaml

streaming: True

# Decoding parameters
# Be sure that the bos and eos index match with the BPEs ones
# Decoding parameters
blank_index: 0
bos_index: 0
eos_index: 0
pad_index: 0
beam_size: 20
nbest: 1
# by default {state,expand}_beam = 2.3 as mention in paper
# https://arxiv.org/abs/1904.02619
state_beam: 2.5
expand_beam: 2.5
lm_weight: 0.1

sample_rate: 16000
n_fft: 512
n_mels: 80
win_length: 32

# Transformer
d_model: 256
joint_dim: 640
nhead: 4
num_encoder_layers: 12
num_decoder_layers: 0
d_ffn: 2048
activation: !name:torch.nn.GELU
output_neurons: 5000
dec_dim: 512

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>
   win_length: !ref <win_length>
    
CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
   input_shape: (8, 10, 80)
   num_blocks: 2
   num_layers_per_block: 1
   out_channels: (64, 32)
   kernel_sizes: (3, 3)
   strides: (2, 2)
   residuals: (False, False)

transformer: !new:speechbrain.lobes.models.transformer.TransformerASR.TransformerASR # yamllint disable-line rule:line-length
   input_size: 640
   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>
   activation: !ref <activation>
   encoder_module: conformer
   attention_type: RelPosMHAXL
   normalize_before: True
   causal: False

# We must call an encoder wrapper so the decoder isn't run (we don't have any)
enc: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
   transformer: !ref <transformer>

# For MTL CTC over the encoder
proj_ctc: !new:speechbrain.nnet.linear.Linear
   input_size: !ref <joint_dim>
   n_neurons: !ref <output_neurons>

# Define some projection layers to make sure that enc and dec
# output dim are the same before joining
proj_enc: !new:speechbrain.nnet.linear.Linear
   input_size: !ref <d_model>
   n_neurons: !ref <joint_dim>
   bias: False

proj_dec: !new:speechbrain.nnet.linear.Linear
   input_size: !ref <dec_dim>
   n_neurons: !ref <joint_dim>
   bias: False

emb: !new:speechbrain.nnet.embedding.Embedding
   num_embeddings: !ref <output_neurons>
   consider_as_one_hot: True
   blank_id: !ref <blank_index>

dec: !new:speechbrain.nnet.RNN.LSTM
   input_shape: [null, null, !ref <output_neurons> - 1]
   hidden_size: !ref <dec_dim>
   num_layers: 1
   re_init: True

Tjoint: !new:speechbrain.nnet.transducer.transducer_joint.Transducer_joint
   joint: sum # joint [sum | concat]
   nonlinearity: !ref <activation>

transducer_lin: !new:speechbrain.nnet.linear.Linear
   input_size: !ref <joint_dim>
   n_neurons: !ref <output_neurons>
   bias: False

asr_model: !new:torch.nn.ModuleList
   - [!ref <CNN>, !ref <enc>, !ref <emb>, !ref <dec>, !ref <proj_enc>, !ref <proj_dec>, !ref <proj_ctc>, !ref <transducer_lin>]

Beamsearcher: !new:speechbrain.decoders.transducer.TransducerBeamSearcher
   decode_network_lst: !new:torch.nn.ModuleList 
      - [!ref <emb>, !ref <dec>, !ref <proj_dec>]
   tjoint: !ref <Tjoint>
   classifier_network: !new:torch.nn.ModuleList 
      -  [!ref <transducer_lin>]
   blank_id: !ref <blank_index>
   beam_size: !ref <beam_size>
   nbest: !ref <nbest>
   lm_module: !ref <lm_model>
   lm_weight: !ref <lm_weight>
   state_beam: !ref <state_beam>
   expand_beam: !ref <expand_beam>

Greedysearcher: !new:speechbrain.decoders.transducer.TransducerBeamSearcher
   decode_network_lst: [!ref <emb>, !ref <dec>, !ref <proj_dec>]
   tjoint: !ref <Tjoint>
   classifier_network: [!ref <transducer_lin>]
   blank_id: !ref <blank_index>
   beam_size: 1
   nbest: 1

lm_model: !new:speechbrain.lobes.models.RNNLM.RNNLM
   output_neurons: !ref <output_neurons>
   embedding_dim: 256
   activation: !name:torch.nn.LeakyReLU
   dropout: 0.0
   rnn_layers: 6
   rnn_neurons: 512
   dnn_blocks: 1
   dnn_neurons: 256
   return_hidden: True  # For inference

modules:
   CNN: !ref <CNN>
   enc: !ref <enc>
   emb: !ref <emb>
   dec: !ref <dec>
   Tjoint: !ref <Tjoint>
   transducer_lin: !ref <transducer_lin>
   normalize: !ref <normalizer>
   proj_ctc: !ref <proj_ctc>
   proj_dec: !ref <proj_dec>
   proj_enc: !ref <proj_enc>

tokenizer: !new:sentencepiece.SentencePieceProcessor

# We compose the inference (encoder) pipeline.
encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
    input_shape: [null, null]
    compute_features: !ref <compute_features>
    normalize: !ref <normalizer>
    cnn: !ref <CNN>
    transformer_encoder: !ref <enc>
    proj_enc: !ref <proj_enc>

pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
    loadables:
        asr: !ref <asr_model>
        lm: !ref <lm_model>
        tokenizer: !ref <tokenizer>

# inference stuff

make_tokenizer_streaming_context: !name:speechbrain.tokenizers.SentencePiece.SentencePieceDecoderStreamingContext
tokenizer_decode_streaming: !name:speechbrain.tokenizers.SentencePiece.spm_decode_preserve_leading_space

make_decoder_streaming_context: !name:speechbrain.decoders.transducer.TransducerGreedySearcherStreamingContext # default constructor
decoding_function: !name:speechbrain.decoders.transducer.TransducerBeamSearcher.transducer_greedy_decode_streaming
   - !ref <Greedysearcher>  # self

fea_streaming_extractor: !new:speechbrain.lobes.features.StreamingFeatureWrapper
   module: !new:speechbrain.nnet.containers.LengthsCapableSequential
      - !ref <compute_features>
      - !ref <normalizer>
      - !ref <CNN>
   # don't consider normalization as part of the input filter chain.
   # normalization will operate at chunk level, which mismatches training
   # somewhat, but does not appear to result in noticeable degradation.
   properties: !apply:speechbrain.utils.filter_analysis.stack_filter_properties
      - [!ref <compute_features>, !ref <CNN>]