3.3 Changelog#

3.3.4#

  • Fixed a bug with cutoff modeling for small datasets

  • Added command to Find OOVs in a corpus (mfa find_oovs)

  • Fixed a bug in mfa transcribe_whisper where --vad flag was not creating TextGrid output with the segments generated from VAD

  • Added validation check for corpus directories to prevent overwriting of files in the MFA_ROOT_DIR temporary directory

  • Updated phone duration mean and standard deviation calculation

3.3.3#

  • Fixed a data type mismatch issue when using whisperx with CPU

  • Fixed a bug where files were being multiply transcribed when using torch models on CPU

  • Fixed a python version compatibility issue with paths being saved in configuration files

  • Fixed a compatibility issue with sklearn 1.5

3.3.2#

  • Clarify error message when modules removed in Python 3.13 are not found.

3.3.1#

  • Added better error handling for multiprocessing functions

  • Added more consistent handling of reference phone and word intervals

3.3.0#

  • Added support for incorporating existing reference alignments into acoustic model training and adaptation

  • Added ReferencePhoneIntervals and ReferenceWordIntervals tables separate from those generated from alignments

  • Removed dependency on Biopython’s pairwise2 module for evaluating alignments

  • Added signal-to-noise ratio calculation as part of alignment evaluation

  • Added filters in training and adaptation based on alignment evaluation metrics

  • Modified temporary files to use dictionary names instead of dictionary ids for easier debugging with multiple dictionaries

  • Compatibility with Kalpy 0.7.0

On this page

This Page