Ivector Configuration

For the Kaldi recipe that ivector extractor training is based on, see https://github.com/kaldi-asr/kaldi/blob/master/egs/wsj/s5/steps/online/nnet2/train_diag_ubm.sh and https://github.com/kaldi-asr/kaldi/blob/master/egs/wsj/s5/steps/online/nnet2/train_ivector_extractor.sh

Parameter Default value Notes
ubm_num_iterations 4 Number of iterations for training UBM
ubm_num_gselect 30 Number of Gaussian-selection indices to use while training
ubm_num_frames 500000 Number of frames to keep in memory for initialization
ubm_num_gaussians 256  
ubm_num_iterations_init 20 Number of iteration to use when initializing UBM
ubm_initial_gaussian_proportion 0.5 Start with half the target number of Gaussians
ubm_min_gaussian_weight 0.0001  
ubm_remove_low_count_gaussians True  
ivector_dimension 128 Dimension of extracted ivectors
num_iterations 10 Number of training iterations
num_gselect 20 Gaussian-selection using diagonal model: number of Gaussians to select
posterior_scale 1.0 Scale on posteriors to correct for inter-frame correlation
silence_weight 0.0  
min_post 0.025 Minimum posterior to use (posteriors below this are pruned out)
num_samples_for_weights 3  
gaussian_min_count 100  
subsample 5 Speeds up training (samples every Xth frame)
max_count 100  
apply_cmn True Flag for whether to apply CMVN to input features

Default training config file

  type: "mfcc"
  use_energy: true
  frame_shift: 10

  - ivector:
      num_iterations: 10
      gaussian_min_count: 2
      silence_weight: 0.0
      posterior_scale: 0.1
      max_count: 100