AcousticModelTrainingMixin#

class montreal_forced_aligner.acoustic_modeling.base.AcousticModelTrainingMixin(identifier, worker, num_iterations=40, subset=0, max_gaussians=1000, boost_silence=1.0, power=0.25, initial_gaussians=0, optional=False, **kwargs)[source]#

Bases: AlignMixin, TrainerMixin, FeatureConfigMixin, MfaWorker, ModelExporterMixin

Base trainer class for training acoustic models and ivector extractors

Parameters:

identifier (str) – Identifier for the trainer
worker (AcousticCorpusPronunciationMixin) – Top-level worker
num_iterations (int) – Number of iterations, defaults to 40
subset (int) – Number of utterances to use, defaults to 0 which will use the whole corpus
max_gaussians (int) – Total number of gaussians, defaults to 1000
boost_silence (float) – Factor by which to boost silence during alignment, defaults to 1.25
power (float) – Exponent for number of gaussians according to occurrence counts, defaults to 0.25
initial_gaussians (int) – Initial number of gaussians, defaults to 0

See also

AccStatsFunction: Multiprocessing helper function for each job
AcousticModelTrainingMixin.acc_stats_arguments: Job method for generating arguments for the helper function
gmmbin/gmm-sum-accs.cc: Relevant Kaldi binary
gmmbin/gmm-est.cc: Relevant Kaldi binary
train_mono.sh: Reference Kaldi script
train_deltas.sh: Reference Kaldi script

acc_stats_arguments()[source]#

Generate Job arguments for AccStatsFunction

Returns:: Arguments for processing
Return type:: list[AccStatsArguments]

acoustic_model_training_params()[source]#: Configuration parameters

align_iteration()[source]#: Run alignment for a training iteration

property alignment_model_path#: Alignment model path

abstract compute_calculated_properties()[source]#: Compute any calculated properties such as alignment iterations

property corpus_output_directory#: Directory of the corpus

property data_directory#: Get the current data directory based on subset

property db_engine#: Top-level worker’s database engine

property db_string#: Root worker’s database connection string

export_model(output_model_path)[source]#

Export an acoustic model to the specified path

Parameters:: output_model_path (str) – Path to save acoustic model

property exported_model_path#: Model path to export to once training is complete

finalize_training()[source]#: Finalize the training, renaming all final iteration model files as “final”, and exporting the model to be used in the next round alignment

property gaussian_increment#: Amount by which gaussians should be increased each iteration

increment_gaussians()[source]#: Increment the current number of gaussians

initialize_training()[source]#: Initialize training

property jobs#: Top-level worker’s job objects

property meta#: Generate metadata for the acoustic model that was trained

property model_path#: Current acoustic model path

property next_model_path#: Next iteration’s acoustic model path

property num_current_utterances#: Number of utterances of the corpus

property phone_type#: Phone type, not implemented for BaseTrainer

property previous_aligner#: Previous aligner seeding training

session(**kwargs)[source]#: Top-level worker’s database session

train()[source]#

Train the model

Raises:: KaldiProcessingError – If there were any errors in running Kaldi binaries

train_iteration()[source]#: Perform an iteration of training

property train_type#: Training type, not implemented for BaseTrainer

utterances(session=None)[source]#

Get all utterances in the trainer’s root worker

Parameters:: session (sqlalchemy.orm.Session, optional) – Session to use in querying
Returns:: Utterance query
Return type:: sqlalchemy.orm.Query

property working_directory#: Training directory

property working_log_directory#: Training log directory