TriphoneFmllrConfig

class aligner.config.TriphoneFmllrConfig(align_often=True, **kwargs)[source]

Configuration class for speaker-adapted triphone training

Scale options defaults to:

['--transition-scale=1.0', '--acoustic-scale=0.1', '--self-loop-scale=0.1']

If align_often is True in the keyword arguments, realign_iters will be:

[1, 5, 10, 15, 20, 25, 30, 35, 38]

Otherwise, realign_iters will be:

[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 16, 18, 20, 23, 26, 29, 32, 35, 38]

fmllr_iters defaults to:

[2, 4, 6, 12]

Attributes

num_iters (int) Number of training iterations to perform, defaults to 35
scale_opts (list) Options for specifying scaling in alignment
beam (int) Default beam width for alignment, defaults = 10
retry_beam (int) Beam width to fall back on if no alignment is produced, defaults to 40
max_iter_inc (int) Last iter to increase #Gauss on, defaults to 30
totgauss (int) Total number of gaussians, defaults to 1000
boost_silence (float) Factor by which to boost silence likelihoods in alignment, defaults to 1.0
realign_iters (list) List of iterations to perform alignment
stage (int) Not used
power (float) Exponent for number of gaussians according to occurrence counts, defaults to 0.25
do_fmllr (bool) Specifies whether to do speaker adaptation, defaults to True
num_states (int) Number of states in the decision tree, defaults to 3100
num_gauss (int) Number of gaussians in the decision tree, defaults to 50000
cluster_threshold (int) For build-tree control final bottom-up clustering of leaves, defaults to 100
fmllr_update_type (str) Type of fMLLR estimation, defaults to 'full'
fmllr_iters (list) List of iterations to perform fMLLR estimation
fmllr_power (float) Defaults to 0.2
silence_weight (float) Weight on silence in fMLLR estimation

Attributes

inc_gauss_count
max_iter_inc

Methods