.. _compare_alignments:

Comparing and evaluating alignments ``(mfa compare_alignments)``
================================================================

As of MFA 3.4, you can compare arbitrary sets of alignments for the same files.  The "reference" directory can be hand corrected or manually segmented files from a benchmark corpus such as TIMIT, Buckeye, Seoul Corpus, or the Corpus of Spontaneous Japanese, which can then be interpreted as boundary errors or boundary precision for benchmarking purposes.  The alignments do not have to have been generated by MFA for this command (unlike when using the ``--reference_directory`` flag for :ref:`pretrained_alignment`).  If the two sets to be compared are both automatically generated, then any boundary errors should be interpreted more as agreement in boundary placement than as a measure of accuracy or evaluation.

If the reference directory specified has audio files or the ``--audio_directory`` parameter is specified, then alignment analysis metrics related to the audio file can be calculated, such as :ref:`alignment_analysis_snr` and :ref:`alignment_analysis_intensity_deviation` of phone intervals.  Without this, other alignment analysis metrics that just rely on timing information such as the :ref:`alignment_analysis_duration_deviation` will still be calculated.  However, since the input is just the alignment outputs and not anything involving an MFA model, metrics involving :ref:`alignment_analysis_log_likelihood` from a model cannot be calculated.

.. seealso::

   * :ref:`alignment_evaluation` for details on the evaluation algorithm.
   * :ref:`alignment_analysis` for details on the fields generated in the ``alignment_analysis.csv`` file in the output folder


Command reference
-----------------

.. click:: montreal_forced_aligner.command_line.compare_alignments:compare_alignments_cli
   :prog: mfa compare_alignments
   :nested: full