hyperparams.yaml

# ################################
# Model: Transducer ASR
# Augmentation: SpecAugment
# Authors: Pooneh Mousavi 2023
# ################################
# Feature parameters (FBANKS etc)
sample_rate: 16000
n_fft: 400
n_mels: 80

# Model parameters
activation: !name:torch.nn.LeakyReLU
dropout: 0.15
cnn_blocks: 3
cnn_channels: (128, 200, 256)
inter_layer_pooling_size: (2, 2, 2)
cnn_kernelsize: (3, 3)
time_pooling_size: 4
rnn_class: !name:speechbrain.nnet.RNN.LSTM
rnn_layers: 5
rnn_neurons: 1024
rnn_bidirectional: True
dnn_blocks: 2
dnn_neurons: 1024
dec_neurons: 1024
joint_dim: 1024

# Outputs
output_neurons: 1000  # BPE size, index(blank/eos/bos) = 0
# transducer_beam_search : True
# Decoding parameters
# Be sure that the bos and eos index match with the BPEs ones
blank_index: 0
bos_index: 0
eos_index: 0

min_decode_ratio: 0.0
max_decode_ratio: 1.0
beam_size: 4
nbest: 1
# by default {state,expand}_beam = 2.3 as mention in paper
# https://arxiv.org/abs/1904.02619
state_beam: 2.3
expand_beam: 2.3
transducer_beam_search: True


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>

enc: !new:speechbrain.lobes.models.CRDNN.CRDNN
    input_shape: [null, null, !ref <n_mels>]
    activation: !ref <activation>
    dropout: !ref <dropout>
    cnn_blocks: !ref <cnn_blocks>
    cnn_channels: !ref <cnn_channels>
    cnn_kernelsize: !ref <cnn_kernelsize>
    inter_layer_pooling_size: !ref <inter_layer_pooling_size>
    time_pooling: True
    using_2d_pooling: False
    time_pooling_size: !ref <time_pooling_size>
    rnn_class: !ref <rnn_class>
    rnn_layers: !ref <rnn_layers>
    rnn_neurons: !ref <rnn_neurons>
    rnn_bidirectional: !ref <rnn_bidirectional>
    rnn_re_init: True
    dnn_blocks: !ref <dnn_blocks>
    dnn_neurons: !ref <dnn_neurons>

enc_lin: !new:speechbrain.nnet.linear.Linear
   input_size: !ref <dnn_neurons>
   n_neurons: !ref <joint_dim>

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.GRU
   input_shape: [null, null, !ref <output_neurons> - 1]
   hidden_size: !ref <dec_neurons>
   num_layers: 1
   re_init: True

# For MTL with LM over the decoder
dec_lin: !new:speechbrain.nnet.linear.Linear
   input_size: !ref <dec_neurons>
   n_neurons: !ref <joint_dim>
   bias: False

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

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

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



tokenizer: !new:sentencepiece.SentencePieceProcessor
# We compose the inference (encoder) pipeline.
encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
    input_shape: [null, null, !ref <n_mels>]
    compute_features: !ref <compute_features>
    normalize: !ref <normalizer>
    model: !ref <enc>

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

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

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