SatTrainer

class montreal_forced_aligner.acoustic_modeling.SatTrainer(subset=10000, num_leaves=2500, max_gaussians=15000, power=0.2, boost_silence=1.0, quick=False, **kwargs)

Bases: TriphoneTrainer

Speaker adapted trainer (SAT), inherits from TriphoneTrainer

Parameters:

• subset (int) – Number of utterances to use, defaults to 10000

• num_leaves (int) – Number of states in the decision tree, defaults to 2500

• max_gaussians (int) – Number of gaussians in the decision tree, defaults to 15000

• power (float) – Exponent for number of gaussians according to occurrence counts, defaults to 0.2

See also

TriphoneTrainer

For acoustic model training parsing parameters

Variables:

fmllr_iterations (list) – List of iterations to perform fMLLR calculation

acc_stats_two_feats_arguments()

Generate Job arguments for AccStatsTwoFeatsFunction

Returns:

Arguments for processing

Return type:

list[AccStatsTwoFeatsArguments]

property alignment_model_path

Alignment model path

calc_fmllr()

Calculate fMLLR transforms for the current iteration

compute_calculated_properties()

Generate realignment iterations, initial gaussians, and fMLLR iterations based on configuration

create_align_model()

Create alignment model for speaker-adapted training that will use speaker-independent features in later aligning.

See also

AccStatsTwoFeatsFunction

Multiprocessing helper function for each job

SatTrainer.acc_stats_two_feats_arguments

Job method for generating arguments for the helper function

gmmbin/gmm-est.cc

Relevant Kaldi binary

gmmbin/gmm-sum-accs.cc

Relevant Kaldi binary

train_sat.sh

Reference Kaldi script

finalize_training()

Finalize training and create a speaker independent model for initial alignment

Raises:

KaldiProcessingError – If there were any errors in running Kaldi binaries

train_iteration()

Run a single training iteration