hyperparams.yaml

# ############################################################################
# Model:
# Author:
# ############################################################################

# Feature parameters
sample_rate: 16000
time_resolution: 0.01 # in seconds (e.g,, 0.01 = 10 ms)
n_fft: 400
n_mels_vad: 40
batch_size: 512

# VAD parameters
cnn1_channels: 16
cnn2_channels: 32
cnn_kernelsize: (3, 3)
rnn_layers: 2
rnn_neurons: 32
rnn_bidirectional: True
dnn_blocks: 1
dnn_neurons: 16
output_neurons_vad: 1

# ECAPA_TDNN
n_mels_ecapa: 80
out_neurons_ecapa: 7205
emb_dim: 192

dataloader_opts:
  batch_size: !ref <batch_size>

# VAD objects
compute_fbank_vad: !new:speechbrain.lobes.features.Fbank
  sample_rate: !ref <sample_rate>
  n_fft: !ref <n_fft>
  n_mels: !ref <n_mels_vad>
  hop_length: !ref <time_resolution> * 1000 # in ms

mean_var_norm_vad: !new:speechbrain.processing.features.InputNormalization
  norm_type: sentence

cnn: !new:speechbrain.nnet.containers.Sequential
  input_shape: [null, null, !ref <n_mels_vad>]
  norm1: !name:speechbrain.nnet.normalization.LayerNorm
  cnn1: !name:speechbrain.lobes.models.CRDNN.CNN_Block
    channels: !ref <cnn1_channels>
    kernel_size: !ref <cnn_kernelsize>
  cnn2: !name:speechbrain.lobes.models.CRDNN.CNN_Block
    channels: !ref <cnn2_channels>
    kernel_size: !ref <cnn_kernelsize>

rnn: !new:speechbrain.nnet.RNN.GRU
  input_shape: [null, null, 320]
  hidden_size: !ref <rnn_neurons>
  num_layers: !ref <rnn_layers>
  bidirectional: !ref <rnn_bidirectional>

dnn: !new:speechbrain.nnet.containers.Sequential
  input_shape: [null, null, !ref <rnn_neurons> * 2]
  dnn1: !name:speechbrain.lobes.models.CRDNN.DNN_Block
    neurons: !ref <dnn_neurons>
  dnn2: !name:speechbrain.lobes.models.CRDNN.DNN_Block
    neurons: !ref <dnn_neurons>
  lin: !name:speechbrain.nnet.linear.Linear
    n_neurons: !ref <output_neurons_vad>
    bias: False
##########################################################

# ECAPA_TDNN objects
compute_fbank_ecapa: !new:speechbrain.lobes.features.Fbank
  n_mels: !ref <n_mels_ecapa>

mean_var_norm_ecapa: !new:speechbrain.processing.features.InputNormalization
  norm_type: sentence
  std_norm: False

embedding_model: !new:speechbrain.lobes.models.ECAPA_TDNN.ECAPA_TDNN
  input_size: !ref <n_mels_ecapa>
  channels: [1024, 1024, 1024, 1024, 3072]
  kernel_sizes: [5, 3, 3, 3, 1]
  dilations: [1, 2, 3, 4, 1]
  attention_channels: 128
  lin_neurons: 192

mean_var_norm_emb: !new:speechbrain.processing.features.InputNormalization
  norm_type: global
  std_norm: False
#####################

vad: !new:torch.nn.ModuleList
  - [!ref <cnn>, !ref <rnn>, !ref <dnn>]

#####################
modules:
  compute_fbank_vad: !ref <compute_fbank_vad>
  compute_fbank_ecapa: !ref <compute_fbank_ecapa>
  cnn: !ref <cnn>
  rnn: !ref <rnn>
  dnn: !ref <dnn>
  mean_var_norm_vad: !ref <mean_var_norm_vad>
  mean_var_norm_ecapa: !ref <mean_var_norm_ecapa>
  embedding_model: !ref <embedding_model>
  mean_var_norm_emb: !ref <mean_var_norm_emb>

pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
  loadables:
    vad: !ref <vad>
    embedding_model: !ref <embedding_model>
    mean_var_norm_vad: !ref <mean_var_norm_vad>
    mean_var_norm_emb: !ref <mean_var_norm_emb>