Acoustic models#
Acoustic models contain information about how phones are pronounced, trained over large (and not-so-large) corpora of speech. Currently only GMM-HMM style acoustic models are supported, which are generally good enough for alignment, but nowhere near state of the art for transcription.
Note
As part of the training procedure, alignments are generated, and so can be exported at the end (the same as training an acoustic model and then using it with the PretrainedAligner. See export_files for the method and Train a new acoustic model (mfa train) for the command line function.
|
Class for storing acoustic models in MFA, exported as zip files containing the necessary Kaldi files to be reused |